Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
FFSSAA IInnssttiittuuttee DDiissccuussssiioonn PPaappeerr SSeerriieess
Financial Research Center (FSA Institute) Financial Services Agency
Government of Japan 3-2-1 Kasumigaseki, Chiyoda-ku, Tokyo 100-8967, Japan
You can download this and other papers at the FSA Institute’s Web site:
http://www.fsa.go.jp/frtc/english/index.html
Do not reprint or reproduce without permission.
Determinants of the CDS
Spreads of Japanese Firms
Before and After the Global
Financial Crisis
Koichi Iwai
DP 2011-2
August, 2011
The views expressed in this paper are those of the authors and do not necessarily reflect the views of the
Financial Services Agency or the FSA Institute.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 1 -
Determinants of the CDS Spreads of Japanese Firms
Before and After the Global Financial Crisis
Koichi IWAI*
Abstract
The global financial crisis has led to even greater interest in credit default swap (CDS) markets. CDS
markets are regarded as one of the underlying causes in the financial crisis, and given the circumstances
that regulatory reviews have been progressed, rigorously examining price formation in CDS markets
before and after the crisis would be important for considering the functions of the CDS market and how
future regulation ought to be. Nevertheless, only a limited understanding of price formation in the
Japanese CDS market has been accumulated so far. The aim of this paper is to empirically verify the
determinants of CDS spreads whose reference entities are Japanese non-financial corporations. I regard
as the possible determinants of CDS spreads both the structural variables and some state variables which
capture overall market trends and macroeconomic development. Analyses in this paper succeed to find
several characteristics of price formation in the CDS market, including some phenomena never reported
before. First, structural variables satisfy the sign condition and are generally significant, but they cannot
fully explain the fluctuation of CDS spreads. Even if market- and macro-related variables are included
in addition to the structural variables as the independent variables, the explanatory power of the model
is generally low. It is arguable that the so-called “credit spread puzzle” exists in the Japanese CDS
market as well. Second, a phenomenon not reported in foreign countries is confirmed, namely, that the
explanatory power of the regression model improves after the financial crisis. A possible reason for this
is that, after the onset of the financial crisis, market fluctuations in foreign countries begin to have a
considerable impact on the Japanese CDS market. Third, it seems that some unobservable systemic
factors become main factors for CDS spreads variations after the financial crisis. These findings will
help to understand the price formation mechanisms in the Japanese CDS market.
Keywords: Credit default swaps (CDSs), Structural model, Dynamic Heterogeneous Panel Model
* Research Fellow, FSA Institute, Financial Services Agency. I would like to acknowledge Professor Akira Kohsaka of
Osaka University and Professor Naoyuki Yoshino of Keio University for the valuable feedback they provided on the
earlier drafts of this paper. This paper represents the personal opinion of the author and is not the official view of the
Financial Services Agency or FSA Institute. Any possible errors in the paper are entirely the responsibility of the author.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 2 -
1. Introduction
The appraisal of credit default swaps (CDSs) changed considerably with the onset of the global
financial crisis in 2007. Prior to the financial crisis, it seemed to be blindly believed that CDS markets
lead to improvements in social welfare through improving the efficiency of resource allocation and
lessening the financial restrictions on households and business. Since various transactions, including
short selling, could be conducted with very small transaction costs in the CDS markets, investors with
diverse trading motives, such as hedging or speculation, are supposed to participate in the markets,
which could lead to more active information production and more efficient price discovery on the
reference entities. Nevertheless, after the onset of the financial crisis, attention began to shift to negative
views, including that CDS markets had amplified the systemic risk of the entire financial system and
that they had lowered social welfare by provoking moral hazard and excessive risk taking in the banking
sector. Furthermore, financial authorities show concern over unfair trading in CDS markets whereas
several incidents of insider trading extending into CDS markets and stock markets occurred in some
countries. Regulators in advanced countries began to examine measures for tightening the regulation
against CDS markets, and some of them have already been implemented.1
Amid rising concern over CDS markets, empirical examinations on the functions and price
formation in U.S. and European CDS markets have been advanced. The determinants of CDS spreads
and the price discovery among CDS markets, bond markets, stock markets, and other related markets
are designated as key research topics. In contrast, as for the Japanese CDS market, empirical
investigations have been limited so far, so the characteristics of price formation are little understood. In
particular, no empirical research has been reported on the CDS spreads of non-financial firms in Japan.
Unlike CDSs in which a financial institution or the Japanese government is the reference entity, most
CDS transactions of non-financial corporations are conducted domestically. Therefore, it is preferable
that non-financial corporations are included in assessing price formation in the Japanese CDS market. In
this paper, I will focus my attention on Japanese non-financial corporations, and will examine the
determinants of their CDS spreads and the characteristics of price formation before and after the
financial crisis.
The rest of this paper is structured as follows. In Chapter 2, I review previous studies and summarize
the main aims of this paper. Chapter 3 explains the empirical method used in this paper. As the possible
determinants of CDS spreads, I incorporate not only the structural variables but also a number of state
variables, such as interest rates and others that indicate stock market trends. Chapter 4 shows the results
of empirical analyses. I examine the validity of the structural model and discuss the significance of the
1 Positive views about CDS markets can be found in Hellwig (1994), Allen and Gale (2006), Hakenes and Schnabel
(2008), and Aschcraft and Santos (2009). Conversely, negative views have been detailed by Partnoy and Skeel (2007),
Brunnermeier (2008), Hellwig (2008), and Hakenes and Schnabel (2009). For the regulations pertaining to CDS markets
and details of recent changes, refer to BIS (2008) and IOSCO (2009).
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 3 -
state variables. I also turn my attention to the differences and characteristics before and after the recent
financial crisis. In Chapter 5, I draw my conclusions.
2. Previous literature
One of the widely accepted theoretical models on the determinants of credit risk is known as the
structural model or structural approach.2 Structural models usually define a firm’s value as a stochastic
process, regard the situation where the firm value falls below a certain threshold (debt value) as default,
and calculate the frequency (probability) of such a situation occurring based on option pricing theories.
In a simplest structural model, the credit risk of a firm can be determined by three structural variables:
leverage, volatility of firm values and the risk-free interest rate. Credit risk can be expressed as an
increasing function of leverage and the volatility, and a decreasing function of risk-free interest rate.3 A
number of empirical studies have been carried out with regard to this theoretical conclusion: To what
extent can the levels and fluctuations of credit spreads be explained by the three structural variables? If
structural variables cannot fully explain the variation of credit spreads, what mechanisms are driving the
credit spreads’ fluctuation?
While a number of previous studies on bond spreads have confirmed that the structural variables
satisfy the theoretical sign conditions, it has also been reported that the explanatory power of the
structural models is generally low. This low explanatory power is often called the “credit spread puzzle.”
For example, Collin-Dufresne et al. (2001) investigate the determinants of bond spreads in the U.S.
using a linear model in which structural variables are the explanatory variables, and report that the R2
remains generally around 0.2 to 0.3. Faced with the credit spread puzzle, researchers have focused on
other factors that might affect credit risk besides structural variables, such as taxes, liquidity risk and
other systemic risks.4 However, the mechanisms and variables by which fluctuations in credit risk can
be explained with a sufficiently high degree of accuracy are yet to be discovered at the present point in
time.
Recently, there has been another stream of studies examining the determinants of credit risk, that is,
analyses of CDS markets. In evaluating credit risk, CDS spreads could be better indicators than bond
spreads for a number of reasons.5 Longstaff et al. (2005) report that the majority of CDS spreads can be
2 Merton (1974), Black and Cox (1976), Longstaff and Schwarz (1995), Collin-Dufresne and Goldstein (2001), etc.
Ammann (2001) describes various structural models. 3 Depending on the assumptions of the model, it is possible that a positive relationship can occur between the risk-free
interest rate and the credit risk (see, Longstaff and Schwartz (1995) for example). However, our discussion will consider
that the structural model “assumes a negative relationship between the risk-free interest rate and the credit risk” for
simplicity of explanation,. 4 See, Elton et al. (2001) and Driessen (2005), for example. 5 Two main reasons have been suggested. First is that CDS spreads are less arbitrary than bond spreads: the level of bond
spreads changes depending on what is used for the benchmark (risk-free interest rate) while CDS spreads are calculated
as the premium for a notional principal amount and no benchmark-related problems occur. Second, short selling on bond
markets is quite difficult in reality, whereas in the CDS markets short sales can be conducted relatively easily. For this
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 4 -
explained using firm-specific bankruptcy risks. Ericsson et al. (2009) focus on structural models, and
investigate the determinants of the changes of CDS spreads. They report that: (i) the estimated
parameters of the structural variables are consistent with theory; (ii) the lower the credit-rating, the
greater the absolute value of the parameters; (iii) while the R2 is about 0.6 in the case of levels
regression, it is around 0.2 in the case of difference regression. Blanco et al. (2005) also measure the
effects of structural variables on CDS spreads while their focus is on comparing price discovery in the
U.S. bond market and CDS market. They indicate that the explanatory power of the model measured by
R2 is low, at about 0.25, while signs of coefficients are in accordance with the theory. Di Cesare and
Guazzarotti (2010) study the CDSs of U.S. non-financial firms before and after the global financial
crisis. They direct their attention not only to structural variables but also to various market variables,
such as stock price returns, difference between long-term and short-term interest rates, corporate bond
spreads, the U.S. VIX index, and the theoretical value of CDS spreads. They report: (i) about half of all
fluctuations in CDS spreads can be explained using their explanatory variables, and the explanatory
power is higher than existing research; (ii) significant difference in the explanatory power of the model
before and after the financial crisis cannot be detected; (iii) CDS spreads of many issuers increased
simultaneously during the crisis situation, which implies that some common factors have influence on
the CDS market.
Studies on the determinants of credit risk for Japanese firms include Ito and Harada (2004), Baba
and Inada (2009), Ooyama and Sugimoto (2007), Shirasu and Yonezawa (2007), Inaba (2007),
Nakashima and Saito (2009), and Ooyama and Hongo (2010). Of these, Ooyama and Sugimoto (2007),
Shirasu and Yonezawa (2007), and Nakashima and Saito (2009) are few who analyze non-financial
corporations. Ooyama and Sugimoto (2007) use OLS estimation to examine the determinants of bond
spread changes, and report that: (i) although a negative relationship is often observed between the risk-
free interest rate and bond spread, positive relationships also occur for some issuers; (ii) an increase in
swaption volatility, which is supposed to represent the uncertainty of interest rates in the future,
coincides with the increase in the bond spread; (iii) bond spreads seem to have no clear relationship with
the stock price index return and the volatility thereof; (iv) R2 is at a low level, which implies that some
other factors not incorporated into the estimation model might cause changes in the bond spread.
Shirasu and Yonezawa (2007) take the effects of the economic environment and liquidity factor into
account as possible determinants of bond spread in addition to the intrinsic attributes of individual
issues. Although the focus of their analysis is on investigating the presence of the “flight to quality” and
“flight to liquidity” phenomena, we can confirm the following from the analyses described in the paper:
(i) the signs of coefficients on the structural variables are different to those expected in the structural
model for some firms; (ii) overall stock market trends and bond spreads are inversely correlated; (iii)
contrary to theoretical inference, the difference of the risk-free interest rate has a positive effect on bond
spreads. Nakashima and Saito (2009) propose to examine the determinants of bond spreads based on an
reason, it is expected that the evaluations of credit risk by market participants will be reflected more accurately in CDS
markets.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 5 -
estimation model in which time dummies and other factors are added to the structural model. They
report that (i) estimates of the structural variables are consistent with the structural model, (ii) overall
trends in financial markets affect bond spreads as stock price returns and change of bond spreads have a
significant negative relationship. By comparing these existing studies on bond spreads, we can confirm
that there are differences in the sign and significance of the parameter estimates for structural variables
and other state variables.
There are many studies on the Japanese CDS market, but most of those are about CDSs whose
reference entity is a major financial institution. Ito and Harada (2004) suggest that CDS spreads are an
effective indicator capturing the credit risk of major banks. Inaba (2007) employs a structural model to
examine the determinants of CDS spreads for three major banks. He adopts structural variables,
swaption volatility and other variables as determinants of CDS spreads, and reports that the CDS
spreads of major banks are determined consistently with the structural model and the spreads are also
affected by the macro economy, such as by business trends. Baba and Inada (2009) explore the
determinants of subordinated debt spreads and CDS spreads of four major banks, and further investigate
the price discovery between the two spreads. They report that CDS spreads negatively react to stock
price returns, and are positively related to the Japanese sovereign CDS spreads and the historical
volatility of CDS spreads.
As shown above, previous studies on the credit risk of Japanese firms have been accumulated
primarily around bonds and CDSs of financial institutions. As far as I know, no empirical research that
investigates CDSs of non-financial firms has been reported so far. As mentioned earlier, in view of the
limitations inherent in bond spreads, we should be wary about relying on the results of bond market
analyses alone for examining the determinants of credit risk. Therefore, I will focus on the CDS
transactions of non-financial corporations and examine their determinants. Specifically, I will clarify the
extent to which actual fluctuations of CDS spreads can be accounted for by structural variables, and
then will look at the mechanisms that can explain the part which cannot be captured by structural
variables. Furthermore, I will compare the price formations of before and after the recent financial crisis.
3. Methodology
3.1. Selection of variables
Table 1 shows the list of variables in this study. Although no commonly acknowledged definition of
the credit risk variable has been established in existing research, four different definitions — level value
of credit risk spreads, level difference, logarithmic level value, and logarithmic difference — are often
used.6 Therefore, this paper will discuss the results of empirical analyses using these four definitions —
6 The studies by Ericsson et al. (2009), Ooyama and Hongo (2010), Nakashima and Saito (2009), Duffee (1998), Collin-
Dufresne et al. (2001), Di Cesare and Guazzarotti (2010), Greatrex (2008), Ooyama and Sugimoto (2007), Alexopoulou
et al. (2009), and Gai et al. (2009) utilize level values and the differences thereof. In contrast, Edwards (1984), Das et al.
(2009), and Forte (2009) suggest that logarithmic level values are theoretically better. Inaba (2007) uses logarithmic level
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 6 -
defined as , , , and respectively — in so far as the explanation does
not become complicated.
Explanatory variables are comprised of structural variables and other state variables. The structural
variables are leverage ( ), firm value volatility ( and the risk-free interest rate
( ). In calculating , periodic financial data are transformed to daily data using linear
interpolation.7 Firm value and its volatility cannot be observed directly in market places. Existing
studies often use one of three proxy variables: volatility derived from option pricing theory, historical
volatility, and GARCH-type volatility. In this paper, an option-based volatility measure that is
calculated in accordance with the procedure outlined in the Appendix is selected.8 Furthermore it is not
necessarily evident which interest rate should be used for the risk-free interest rate. Most existing papers
use either the yields on long-term government bonds or long-term swap rates. Similar to Ericsson et al.
(2009) and others, I will use the ten-year government bonds yield as risk-free interest rate ( ).9 As the
treatment in many of previous papers, I add the square of the risk-free interest rate, ( ), as an
explanatory variable. Readers should note that “structural variables” do not include the square of the
risk-free interest rate in the explanation below.
Since the credit spread puzzle was acknowledged, variables expressing the macro environment and
market trends have been adopted as the possible determinants of credit risk. These variables could be
construed as state variables for grasping the term structure of interest rates, the stochastic variation of
firm values, the time-varying recovery rate and so forth. However, because the rigorous mathematical
relation between these variables and credit risk has not been explicitly derived in most theoretical
models, the questions of which indicators to be incorporated into the estimation model and how to
incorporate them have to be left to the discretion of researchers.10
I select the following variables that
have been used frequently in existing studies as independent variables that capture the macro
values in order to stabilize the variance of the dependent variable. Forte and Peña (2009) and Ferrucci (2003) utilize
logarithmic differences. 7 In empirical analyses of credit risk, it is normal to use linearly interpolated financial data. For example, Collin-Dufresne
et al. (2001), Greatrex (2008), Ericsson et al. (2009), Nakashima and Saito (2009), and Ooyama and Hongo (2010) utilize
linear interpolation technique. 8 Analyses are also conducted using the GARCH volatility (based on GARCH (1,1) model) and the historical volatility
(90-day) of stock price returns as the proxy variable for firm value volatility. The outcomes are virtually the same as the
findings in this paper. Thus, the claims made in this paper would also hold true for other proxy indexes for firm value
volatility. 9 Forte and Lovreta (2009) remark that it would be inappropriate to use the swap interest rate as a proxy variable for the
risk-free interest rate during the recent financial crisis situation. 10 For example, with regard to the positioning of state variables in structural models, the leading study examining the
determinants of CDS spreads, Collin-Dufresne et al. (2001), only goes as far as presenting the relational expression CS(t)
= CS(Vt, rt, [Xt]), where CS is credit spread, V is firm value, r is the swap rate, X is a variety of state variables, and t is
time. But recently, using variables that grasp macroeconomic trends and overall market trends in credit risk evaluation
models in an ad hoc fashion has become fairly widespread. At the same time, statistically evaluating the effects of these
variables on credit risk has also been advanced. For example, Sommar and Shahnazarian (2009) take a statistical
approach in examining the effects of industrial production, commodity prices and short-term interest rates on credit risk;
and Simons and Rolwes (2009) take a similar approach in examining the effects of GDP, interest rates, foreign exchange,
equity returns and the fluctuations thereof, as well as crude oil prices on credit risk. Chau-Lau (2006) presents a summary
of statistical attempts that incorporate macro factors in a credit risk model.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 7 -
environment and market trends: , the difference between long-term and short-term interest
rates ( ), return on the domestic stock market ( ), the logarithmic market capitalization
( ), swaption volatility ( ), and U.S. VIX index ( ).11
Additional explanations are necessary for some variables. As discussed in detail by Di Cesare and
Guazzarotti (2010), the difference between long-term and short-term interest rates could have either a
positive or negative effect on credit risk spread. Considering this ambiguity, I regard as a control
variable in this paper. There is also uncertainty surrounding the interpretation of stock market returns.
Blanco et al. (2005) is among the scholars who view a stock market return as a proxy variable for
default recovery rates; but it could also be thought of as a state variable grasping macro environment
trends. Although the return on individual stocks could possibly be added as an explanatory variable, I
have not adopted it in this paper because it could have a high correlation with leverage and market
capitalization. Logarithmic market capitalization is generally used as an indicator for firm size. If firm
size correlates with trading volume in the CDS market, we could also regard as a proxy of
liquidity. Meanwhile, gathering necessary data to evaluate the overall liquidity of the Japanese CDS
market and the liquidity of individual CDS issues is difficult, so I have positioned as a control
variable for liquidity concurrently with firm size. Swaption volatility could be regarded as an indicator
capturing fluctuations of the risk-free interest rate and the uncertainty over future prospects of the
interest rate environment. U.S. VIX index has been construed in previous studies as a proxy that reflects
the uncertainty of the overall U.S. stock market as well as global event risk. No matter which
interpretation is adopted, can be regarded as an instability factor in overseas markets from the
perspective of the Japanese CDS market or Japanese investors.
3.2. Estimation model
Many previous studies use the linear model of equation (1).
K
ktitkikiiti xSpread
1,,,,,
)(1
Spreadi,t is the variable representing the credit risk of firm i at time t, and
are used separately as this variable. xi,k,t is the kth
explanatory variable of firm i at time t, and
the explanatory variables (k=1,2,…,K) are comprised of structural variables and other state variables as
mentioned in the previous section. εi,t is the error term of firm i at time t. Previous research can be
divided broadly depending on whether equation (1) is perceived as a time series model of each
individual firm or it is treated as panel data. Collin-Dufresne et al. (2001), Ericsson et al. (2009) and
Ooyama and Sugimoto (2007) are among those who use the OLS method to estimate equation (1) for
each firm. Inaba (2007) also makes estimates for each individual firm, but uses the maximum likelihood
method in which a moving average (MA) process is assumed in εt. In contrast, Di Cesare and
11 When selecting non-structural variables, I referred to Collin-Dufresne et al. (2001), Blanco et al. (2005), Pan and
Singleton (2006), Inaba (2007), Ooyama and Sugimoto (2007), Greatrex (2008), Ericsson et al. (2009), Nakashima and
Saito (2009), and Di Cesare and Guazzarotti (2010).
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 8 -
Guazzarotti (2010), Baum and Wan (2010) use pooled OLS, fixed-effect models and random-effect
models. In addition, Nakashima and Saito (2009) use an instrumental variable fixed-effect model which
takes into account the endogeneity of credit risk spread and leverage, whereas Ötker-Robe and Podpiera
(2010) use a dynamic panel model. Most studies that use panel analysis impose a constraint that the
parameters excluding constant terms are homogenous for all i. If the true parameters are homogenous,
estimating time series models for each firm will lead to loss in efficiency. On the other hand, it is known
that imposing homogeneity in parameters in the panel model will result in biased estimators in cases
where the parameters are heterogeneous in the true model.12
Using the dynamic heterogeneous panel (DHP) model is one possible solution to these problems.
The DHP model is a model that allows parameters to differ in cross-section. Ferrucci (2003),
Alexopoulou et al. (2009), and Gai et al. (2009) use the DHP model in their empirical analysis of credit
risk. Following these previous attempts, this paper estimates a DHP model whose exact formulation is
derived as follows. First, I suppose that credit spread (Spread) follows the autoregressive distributed lag
model (ARDL (p, q1, …, qk)) as shown in equation (2).
p
j
q
jitijtiijjtiijit XSpreadSpread
1 0,
',
)(2
Xit is a K×1 vector composed of the explanatory variables while δij is the K×1 vector of the parameters
and λij is a constant parameter. μi is the fixed effect specific to i, and εit is the error term. The
explanatory variables include structural variables and other state variables. Reparameterization of
equation (2) will produce the following error-correction representation:
1
1
1
0,
*,
*'1,
p
j
q
jitijtiijjtiijititiiit XSpreadXSpreadSpread
)(3
where
q
jmim
*ij
p
jmim
*ij
q
j
K
kikiji
p
jiji
)q,...,,j(
),p,...,,j(,/,
1
10 11
121
12111
Inside the brackets of equation (3) is an expression that represents the long-run equilibrium. θi is called
the long-run parameter. ϕi is a parameter that captures how fast the divergence is adjusted if the credit
risk spread diverges from the long-run equilibrium level. In this paper, ϕi is called the adjustment
parameter. If the adjustment parameter is negative and significant, the divergence converges toward the
long-run equilibrium even if it occurs in the short run. The second and third terms on the right-hand side
are the parts which reflect the short-term movements of the dependent variable. Throughout this paper,
this is called the short-run adjustment equation. Pesaran et al. (1999) propose a pooled mean group
(PMG) model, in which θi in equation (3) is common to all i, but which permits other parameters to be
heterogeneous, and they derive the maximum likelihood estimator. Pesaran and Smith (1995) and
12 See Pesaran and Smith (1995), Pesaran and Shin (1998), Pesaran et al. (1999), etc.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 9 -
Pesaran et al. (1999) propose another estimator, called the mean group (MG) estimator. In this model, θi
is also thought to differ for each i, and so equation (2) is estimated for each i, and then the mean value
of those estimates is calculated. Selection between the PMG model and MG model can be made by the
Hausman test. In this paper, following Alexopoulou et al. (2009) and Gai et al. (2009), I will focus on
an ARDL(1,1,…,1) model with an imposed lag of one period.13
The DHP model is useful in the analysis of financial asset prices which generally follow the I(1)
process, and could be a suitable method when investigating an equilibrium mechanism. There are four
reasons for this. First, the DHP model will produce a consistent and asymptotic normal estimator even if
stationary variables and the I(1) process are mixed together in the model.14
Second is the fact that the
DHP model explicitly analyzes long-run equilibrium relationships. This means that, if a divergence
from the long-run equilibrium relationship occurs, we can examine whether there is a force working to
restore the equilibrium. Thirdly, using the DHP model, we can also analyze how the dependent variable
responds to changes in explanatory variables in the short term. Moreover, the DHP model allows
different short-term responses for each individual. Fourth, the DHP model can also be applied to data
sets with small N but large T sample.
In view of the above discussion, both a time series model using OLS and a DHP model will be
analyzed in this paper. I will report the results of OLS using the four variables as the dependent variable
— , , , and — in order to directly compare the results with the
previous studies. With regard to the DHP model, I will estimate models in which and
are used as the dependent variable.
4. Empirical results
4.1. Descriptive statistics
The sample in this paper covers the CDS transactions where the reference entities are 45 non-
financial firms whose daily data are continuously available for the period from 1 April 2004 to 30
September 2009 (Table 2). Table 2 also shows the list of the reference entities and credit-rating class of
the companies. The sample consists of several sectors; the manufacturing industry accounts for about
half of the sample, with the remainder being such industry sectors as trading companies, real estate,
utilities, and consumer finance. Viewed by the credit-rating, the sample data are distributed across each
class of rating, with main focus on investment-grades.
Table 3 shows several basis statistics. Looking at the correlation between the CDS level variables
and the explanatory variables, we can confirm some characteristics that are consistent with the
13 To be precise, in this paper, I will use an ARDL (1,…,1) model and a model in which parameter constraints are
imposed on the short-run adjustment equation and long-run equilibrium equation in the ARDL (1,…,1) model. A DHP
model in which parameter constraints are imposed on an ARDL model can be seen in Martinez-Zarzoso and Bengochea-
Morancho (2004), Alexopoulou et al. (2009), and Gai et al. (2009). 14 However, certain conditions need to be satisfied, such as T being adequately long. For further details, see the
references listed in footnote 12.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 10 -
theoretical conclusions of the structural model and with existing empirical results. Furthermore, the
maximum absolute value of correlation coefficients is about 0.6, so a multicollinearity is unlikely to be
a serious concern. Table 3-(3) shows the results of the DF-GLS and PP tests of unit roots. ,
, , , and can be confirmed to be the I(1) process for almost
all firms. The results of the unit root tests show that regression analyses using the level variables may
lead to the problem of spurious regression.
4.2. Time series model
Table 4 and Table 5 show the OLS results using the level variable and the difference variable
respectively. 15
The “coef” and “s.e.” in the Tables are the cross-section mean values of the parameter
estimates and standard errors respectively. In contrast, the “t-value” and “p-value” are the t-value and p-
value pertaining to the constant term from the OLS estimation, in which the dependent variable consist
of the parameter estimates derived from OLS regressions for each firm while the explanatory variable is
a constant term.16
In cases where only structural variables are used as explanatory variables (hereafter “base model”),
in almost all formulations, the results are in line with the structural model, regardless of whether the
variables are level variables or difference variables. In other words, estimated coefficients of
and are positive and significant,17
and the coefficient of is negative and
significant, except for one case. In terms of the validity of models, adj-R2 is less than 0.04 for the model
using the difference variables whereas it is relatively high at more than 0.6 for the model using the level
variables.
In addition to structural variables, I select other explanatory variables that are indicative of market
trends. Analyses are separated into two types; one including and the other without it. The model
without is a formulation that considers domestic factors only. In contrast, the model including
can be regarded as the one which takes overseas factors into account. To begin with, looking at the
estimation results of the level variables, we can confirm the following characteristics. First, similar to
the results of the base model mentioned above, the signs of coefficients for structural variables are
consistent with the structural model. However, coefficients of and are insignificant in a
number of specifications. Second, the coefficients of take an opposite sign depending on whether the
left-hand side variable is logarithmic form or not. As mentioned earlier, although parameters for
15 Following Inaba (2007), I re-estimated the model reported below using the maximum likelihood method, assuming
ARMA (1,1) for the error term. Since mostly similar results are obtained, only the results of the OLS estimation are
reported below. 16 To be precise, the t-value (p-value) is the t-value (p-value) for the null hypothesis that the parameter of the constant
term is zero. This approach of using OLS estimates is a simplified method for understanding whether the each
explanatory variable is significant across cross-section. Same method is used in cases where there are a large number of
cross-sectional samples. Collin-Dufresne et al. (2001) and Blanco et al. (2005), for instance, adopt this method for their
empirical analyses of credit risk spreads. Similar methods will be used when reporting the results of the PMG model. 17 In this paper, the significance of hypothesis testing is basically assessed at the 1% level. Therefore, such expressions as
“is significant (is not significant)” refer to assessments at the 1% level, unless otherwise noted.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 11 -
could theoretically be either positive or negative, the fact that the sign changes depending on the form of
the dependent variable is difficult to understand economically. It is expected that this results might be
due to non-stationarity of the variables. Third, the estimated parameters for are positive and
significant in all cases. If we regard as a proxy variable for recovery rate, this result would mean
that an increase in recovery rate leads to an increase in credit risk. The finding that overall stock market
trends have a positive effect on the credit risk spreads of individual firms can also be found in Ooyama
and Sugimoto (2007) and Pynnӧnen et al. (2004), but the reasoning given in these previous studies has
not necessarily been well defined. Fourth, in the model where is used as the dependent variable,
the coefficients for become negative. This result is different from Ooyama and Sugimoto
(2007) and Inaba (2007). Fifth, the coefficients for are positive, indicating the possibility that the
impact of any uncertainty in the U.S. stock market could reach to the Japanese CDS market. Finally,
adj-R2 are relatively high, at about 0.7 to 0.9.
In contrast, looking at the estimation results for the difference variable being used as a dependent
variable, the following points are noted. First, although the coefficients’ sign of the structural variables
are no different to the results of the level variables, statistical significance of some variables changes;
most of the coefficients of turn to be significant whereas the coefficients of lose
significance in the formulation on which is used as the dependent variable. In the same
formulations, none of the parameters of , , or are significant, suggesting
that interest rate movements do not have any effect on . This finding differs from Nakashima and
Saito (2009) who report a statistically significant relationship between credit risk spreads and interest
rates. The second feature is that intuitively easy-to-understand results are obtained; the coefficients of
and are negative and positive respectively in all cases. Third, the coefficients of
are positive and significant, which is similar to the results for the level variables. Fourth, the adj-
R2 rise no higher than 0.03 to 0.07, which is lower than those reported in previous studies on the U.S.
market. The adj-R2 calculated by Blanco et al. (2005) and Ericsson et al. (2009) using weekly data are
about 0.15 to 0.25. Studies that use monthly data, such as Di Cesare and Guazzarotti (2010), Collin-
Dufresne et al. (2001), and Greatrex (2008), report adj-R2 from 0.3 to 0.5. After re-estimating the model
in Table 5 using monthly data (monthly mean values), the adj-R2 are calculated at around 0.3 to 0.4.
Although the differences in explanatory variables and specifications of the models render it impossible
to make simple comparisons, the above results seem to show that the explanatory power of the models
in this paper is, at best, at a similar level as the empirical results of other countries.
4.3. Dynamic heterogeneous panel (DHP) model
The findings in section 4.2 indicate that regression analyses using level variables as the dependent
variable could lead to the problem of spurious regression, and the estimations based on difference
variables lead to the issue of the poor explanatory power. In contrast, as mentioned earlier, one of the
features of the DHP model is that it does not require variables to be stationary. The DHP model can
incorporate the information reflected in both level and differenced variables through the long-run
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 12 -
equilibrium equation and the short-run adjustment equation. Accordingly, we could overcome the above
problems if we use the DHP model.
To begin with I estimate models in which structural variables are the only independent variables, and
then will move to models including other state variables as additional explanatory variables. The actual
estimation model is equation (3) which is a variation of the ARDL model. Because is an indicator
that captures trends in overseas markets, it would not be regarded as a variable that determines the long-
run equilibrium value of CDS spreads for Japanese corporations. Therefore, I regard it as an exogenous
variable and only consider its effects via the short-run adjustment equation. Table 6 shows the
estimation results. Irrespective of the form of dependent variable, the PMG model is selected as a result
of the Hausman test (1% level). Consequently, I will report only the results of the PMG models.
Before expanding on the features of the estimation results, an explanation needs to be given as to the
figures in the Tables. The figures for the long-run equilibrium equation in the PMG model report the
parameters that are common to all i. In other words, the “coef”, “s.e.”, “z-value”, and “p-value” in the
Tables can be interpreted in the same way as ordinal regression analyses. In contrast, with respect to
each of the variables in the short-run adjustment equation, the parameters are permitted to be different
between individual reference entities. The “coef” in the Tables relating to these heterogeneous
parameters is the cross-sectional average of the estimated coefficients. The reported “s.e.”, “z-value”,
and “p-value” of these variables are the standard error, z-values, and p-values relating to the constant
term in the OLS estimation, in which the parameter estimates of each issue are the dependent variables
while a constant term is the only explanatory variable. “χ(p-value)” are test statistics and associated p-
value of the null hypothesis that the parameters are common to each issue. The adj-R2 in the lower part
of the Tables is the cross-sectional average of adj-R2.
Panel A and Panel B in Table 6 report the results of the PMG models where and
are used as a dependent variable respectively. The signs of coefficients in the long-run equilibrium
equation are, on the whole, consistent with the structural model and previous studies. Meanwhile, three
variables among the explanatory variables in the short-run adjustment equation — , , and
— have not been used much as explanatory variables in existing literature. In addition, it
would be hard to think of any economic significance these variables would have on CDS spreads.
Thereupon, I decide to re-estimate the DHP model with zero restrictions imposed on the parameters of
these three variables, and examine the results in detail below.18
The estimation results of the restricted model are shown in Table 7. I estimate two models just as
with Table 6; one including and the other without . The PMG models are selected as a result of
the Hausman test in all cases, and so only the results of the PMG models are reported. The following
features of the estimation results are worthy of attention. First, the coefficients of the structural variables
18 As we could not observe any great difference between the estimation results in Table 6 and Table 7, whether
restrictions are imposed upon , , has no great bearing on the sign or significance of the parameters
of the other explanatory variables. For this reason, the basic assertions below hold true regardless of whether there are
restrictions or not.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 13 -
both in the long-run equilibrium equation and in the short-run adjustment equation have the same signs
as those suggested by the structural model, and moreover, most of the coefficients are significant.
Consequently, the price formation mechanism suggested by the structural model is considered to be at
work. Second, the coefficients of , , and in the long-run equilibrium equation are
significant in all cases. Furthermore, considering the fact that the adjustment parameter is significantly
negative, there is a strong possibility that an equilibrium relationship has formed between
( ) and the variables contained in the long-run equilibrium equation. This result also suggests
limitations of the structural model, in the sense that the structural variables alone are not enough to
explain the equilibrium levels of CDS spreads. In addition, the speed of adjustment towards a long-run
equilibrium is supposed to be different for each CDS since the null hypothesis that the adjustment
parameters are the same for all issues can be rejected. Third is the finding that the parameters of
are significantly positive. Also, the adj-R2 increase considerably by adding to the explanatory
variables. In the estimation of the time series model, and are significant, but it is not
confirmed that adding these variables to the independent variables would cause a considerable increase
in the adj-R2. These results of the PMG model show that the U.S. market trends and perhaps global
event risks have a strong effect over the Japanese CDS market in the short term. However, the level of
the adj-R2 per se could hardly be called high compared to previous studies, and in this sense I would
have to say that the credit spread puzzle exists in Japan too.
As described in footnote 8, the results thus far generally hold true even if we change the definition of
firm value volatility. I will now, however, confirm the robustness of the results from a different
perspective. Specifically, I will use the distance-to-default indicator ( ) as an explanatory variable in
place of three structural variables. As evident from the process for deriving this indicator outlined in the
Appendix, this measure can be thought of as a variable that aggregates the information contained in
three structural variables.19
If this measure reflects credit risk accurately and investors are acting
rationally, the parameters of should be negative and significant. Looking at the estimation results in
Table 8, in all but one case, the signs of coefficients pertaining to are negative, and in most cases,
they are significant. However, positive and significant results can also be seen in the formulation in
which is used as the dependent variable. The fact that the negative and significant result could
not be obtained consistently throughout might suggest that either the structural model providing the
theoretical basis for does not indicate true credit risk, or that investors are not acting rationally in
accordance with the structural model.20
Although we need to be mindful of this point, the estimation
results using confirm the adjustment parameter to be negative and significant, and so there seems to
be a strong possibility that a long-run equilibrium relationship exists between CDS spreads and
19 See Vassalou and Xing (2004), Byström (2006), Das et al. (2009), Bharath and Shumway (2008), and Du and Suo
(2007). 20 This is a composite hypothesis that commonly occurs when examining market efficiency. It is difficult to clarify which
of the two hypotheses is rejected.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 14 -
explanatory variables. Moreover, we could reconfirm such phenomena as an improved adj-R2 by adding
variable in the model.
4.4. Before and after the financial crisis
4.4.1. DHP model
The recent global financial crisis has resulted in attention being switched to more negative
perceptions of CDS markets than positive views. Some authors, such as Di Cesare and Guazzarotti
(2010) and the IMF (2009), note that price formation in CDS markets has changed after the financial
crisis. In the rest of this section, I examine whether the financial crisis prompted a change in the price
formation in CDS spreads in Japan. Dividing the sample into before and after the financial crisis, I will
examine the differences in estimated coefficients. Following the discussion of Di Cesare and
Guazzarotti (2010), I regard the financial crisis happened July 2007. I designate the period from April 1,
2004 through to June 29, 2007 as the pre-crisis period; and the period from July 1, 2007 through to
September 30, 2009 as the crisis period.21
Different formulations in section 4.3 are also conducted, but
since no major differences are found, only the results on the model that includes will be discussed
for the rest of this section.
Table 9 shows the estimation results. Regardless of whether or is used as a
dependent variable, some common features can be identified. First, the significance of individual
variables is more apparent and the overall fitness of the model is better in the second half. Although
some variables in the long-run equilibrium equation and in the short-run adjustment equation are not
significant in the first half, almost all variables are significant in the second half. Furthermore the adj-R2
in the second half is clearly higher than that in the first half. Second, the possibility of stronger
interconnection between CDS markets and bond markets/stock markets can be observed in the second
half. For example, coefficients on interest rate related variables in the second half clearly differ from
those in the first half. In the first half, coefficients of and in the long-run equilibrium
equation as well as the difference series of these variables in the short-run adjustment equation are not
significant; but in the second half, almost all of these coefficients become significant. In addition, the
coefficient of also becomes negative and significant in the second half. Furthermore, the
estimated parameter of is positive in the first half, but switches to negative in the second half. The
first-half results signify that increases in term spreads lead to an expansion of credit risk spreads. This
phenomenon has also been confirmed in Blanco et al. (2005), but there is no consensus as to the
underlying mechanism. In any case, we could say that these results suggest the possibility that,
following the onset of the financial crisis, CDS markets have become more easily susceptible to trends
in other financial markets and in the overall financial market. Third, whereas the adjustment parameter
21 June 2007 is the period when the failure by a hedge fund under the control of Bear Stearns in subprime loan
investments was viewed with apprehension in the marketplace, and when subprime-related securitized papers began to be
downgraded.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 15 -
is -0.003 to -0.005 in the first half, it declines as far as -0.012 to -0.016 in the second half.22
The fact
that the negative range of the adjustment parameter becomes broader suggests that CDS spreads are
moving toward convergence at the long-run equilibrium level at a greater speed.23
4.4.2. Principal component analysis
In the analyses thus far, it is found that the explanatory power is limited no matter how we formulate
the model, except for the time series model using level variables as the dependent variable, for which
there is another concern of spurious regression. Previous studies on the U.S. and European credit
markets report somewhat paradoxical outcomes; the majority of variation in credit spreads that cannot
be captured through regression analysis ends up being explained using some systemic factors.24
I will
examine whether such a phenomenon can also be observed in the Japanese CDS market. Specifically, I
will use principal component analysis to investigate how much of the CDS spread’s fluctuations can be
explained by unobservable common factors.
Table 10 shows the results of the principal component analysis conducted on the four variables;
, , and two residual series from the full models with or as a
dependent variable. Principal component analysis has been separately conducted on samples before and
after the financial crisis. Looking at Table 10, we can confirm that the contribution ratio of the first
principal component of the original series ( , ) increases in the second half. This
suggests that, some factors common to the sample firms become the main determinants of credit risk
during the second half. A similar trend can also be observed for the first principal component of the
residual series of . In other words, it would seem that systemic factors that cannot be fully
understood using explanatory variables in the full model become the main variation factors in
, especially during the second half. Figure 1 is a scatter diagram, with the eigenvector of the
first principal component calculated from the original series of on the horizontal axis, and
the eigenvector of the first principal component calculated from the residuals of the full model on the
vertical axis. As the figure clearly shows, there is a strong positive correlation between them. This
suggests the possibility that systemic factors included in the original series remain in the residuals
without being fully captured even by regression analysis. In light of these results, it is expected that
some systemic factors other than the explanatory variables adopted in the estimation model are causing
fluctuation of CDS spreads.
4.4.3. Discussion
The abovementioned analyses suggest that the price formation in the CDS markets changed around
the time of the financial crisis. Results of the DHP model confirm that the significance of each
22 Even if the estimation periods are changed in various ways, for the most part, these patterns can be confirmed. 23 This does not necessarily mean that CDS spreads are being determined close to the long-run equilibrium level. In other
words, despite the adjustment speed accelerating, it is feasible that the gap between the CDS spreads and the long-run
equilibrium during the second half period will be broader than that in the first half. 24 See Collin-Dufresne et al. (2001), Ericsson et al. (2009) and Di Cesare and Guazzarotti (2010).
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 16 -
explanatory variable and the overall fitness of the model improved after the financial crisis. These
finding are arguably contrary to the previous studies on the U.S. and European markets.25
What are the
reasons for the explanatory power of the model to improve amid the crisis? The first thing that comes to
mind is the possibility that the Japanese CDS market has begun to be strongly influenced by market
trends in foreign countries, in particular in the U.S. stock market. This is confirmed by the coefficients
of . Pan and Singleton (2006), for instance, also confirm that the U.S. VIX index has a significantly
positive effect on the CDS spreads whose reference entity is the Japanese government or major Japanese
banks. In contrast, to the best of the author’s knowledge, no empirical finding has been reported on the
U.S. and European markets, which shows that adding the VIX index as an independent variable
improves the explanatory power of a model after the onset of the financial crisis. Second, CDS spreads
seem to have been determined with a fair degree of “independence” from the bond and stock markets
prior to the financial crisis. One of the notable features of the analyses before the financial crisis is that
almost all the variables that are related to the other financial market — such as , and —
are not significant while the coefficients of and are significant. This suggests
that the interconnection between CDS market and bond/stock market did not manifest before the
financial crisis. In other words, the onset of the crisis has led to various transactions extending across
these markets, and as a result, the information efficiency of market prices has improved. The changes in
the adjustment parameter seen in Table 9 indirectly support this hypothesis.
On the other hand, the results of the principal component analysis suggest that systemic factors
common to all firms triggered fluctuations of CDS spreads during the financial crisis. The question of
what these systemic factors are is an issue for the future research. However, given that it is getting clear
that liquidity is having a notable impact on price formation in the U.S. CDS markets, and that variables
related to the overall liquidity of CDS markets and to the liquidity of individual CDSs have not been
considered in this paper, liquidity might be related to the systemic factors. As of the time of writing this
paper, there are no public data on both overall liquidity of the Japanese CDS market and liquidity of
individual CDSs. When relevant data become accessible, we will need to confirm the effect of the
liquidity on price formation in the Japanese CDS market.
5. Conclusion
In this paper, I examined the pricing of the Japanese CDSs in which non-financial corporations are
the reference entity. I relied on a simple time series model and DHP model, focusing on structural
variables and other state variables that capture financial market and macroeconomic trends. The results
confirm that price formation mechanisms suggested by the structural model are at work, but that price
fluctuations in the CDS market cannot be fully explained even using structural variables and other
25 For a comparison before and after the financial crisis and the effects of VIX, see Di Cesare and Guazzarotti (2010),
Ferrucci (2003), and Greatrex (2008). For the effects of liquidity, see Driessen (2005), Tang and Yan (2008), Acharya et
al. (2008), Bongaerts et al. (2008, 2011), and Nashikkar et al. (2009).
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 17 -
market variables. I also point out that, price formation in CDS markets changed when the financial crisis
hit, and some kind of systemic factors have been influencing CDS spreads particularly since the onset of
the financial crisis.
In conclusion, I will summarize the implications derived from these results, separated into
suggestions for research and significance for public policy and financial business. With respect to the
suggestions for future research, first, it can be pointed out that using a model like the DHP model, which
allows heterogeneous parameters and long-run equilibrium relationships, is preferable when estimating
CDS spreads. While time series models and panel analyses with homogeneous parameter restrictions
have been widely used in existing studies, the results in this paper show the possible superiority of the
DHP model over the existing quantitative models. It should be noted that in addition to structural
variables, a range of variables that capture market trends need to be incorporated as determinants of
CDS spreads. In particular, given that U.S. VIX has high explanatory power in analyses in this paper,
effects from overseas need to be taken into full account when estimating the determinants of credit risk
in domestic markets. Second, the CDS market seems not to have been linked to the bond and stock
markets for the pre-crisis period. If this low-level interconnection has been caused by the lack of an
arbitrage function extending across these markets, then we should be cautious about the possibility that
the efficient allocation of resources and the risk sharing function through the CDS market have not been
fully manifested. In addition to a more detailed investigation into the interconnection between CDS
markets and other financial markets, the effects of CDS markets on economic welfare will also need to
be examined. On the other hand, the fact that CDS spread fluctuations have not been fully elucidated —
the credit spread puzzle — has important implications for financial businesses. That is, this uncertainty
in pricing will need to be fully considered when conducting CDS trading and associated risk
management. It is also expected that this puzzle will make it difficult for regulatory authorities to assess
the validity of CDS spreads. As a consequence, it may be becoming more difficult to identify unfair
trading related to CDS markets.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 18 -
Appendix: Derivation of Firm Value Volatility and the Distance-to-Default Indicator
In this appendix, I describe the method for deriving firm value volatility ( ) and the
distance-to-default measure ( ). The concept underlying the calculation of and is
the nature of equity as a call option (Black and Scholes 1972, 1973., etc.). In other words, an equity can
be viewed as the right of claim against residual assets in case where the firm value exceeds the debt
value at some point in the future (maturity). The method of derivation below follows Gropp et al. (2002).
A detailed description of the derivation and the characteristics of the DD can be found in the paper, and
so will be omitted here.
In preparation for the explanation, I shall define a number of variables as follows:
VE: market value of equity (market capitalization)
VA: firm value
D: debt value
r: risk-free interest rate
ζA: volatility of a firm’s asset value
ζE: equity volatility
T: time until the maturity of the debt liabilities
ε: standard normal distribution
N():cumulative standard normal distribution function
dz: standard Wiener process
To begin with, if we assume the premises of the basic option pricing theory by Black and Scholes
(1972), the following relationship between firm value and stock value applies.
21 dNDedNVV rTAE
Tdd
T
TrD
V
d
A
A
AA
12
2ln
1
2
)(i
Here, the following equation holds true between firm value volatility and stock volatility:
A
E
AE dN
V
V 1
)(ii
Thus, if VE, ζE, D, r, and T are once determined, we can calculate VA and ζA from equations (i) and (ii)
using convergence calculation. In this paper, VA and ζA are calculated assuming VE is market
capitalization, ζE is the (60-day) historical volatility of the stock price series, D is the book value of the
total debt liabilities, r is the yield on government bonds, and T is 1 (year). The ζA calculated by
following this procedure is the firm value volatility ( ) set forth in this paper.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 19 -
If we consider that firm value follows the geometric Brownian motion, we get the following formula
for firm value at time t.
ttrVV AA
AtA
2lnln
2
)(iii
Here, we regard a default to be the state that firm value is less than its debt value until the maturity of
the debt. In this setting, the default probability can be defined by equation (iv), and we can define the
distance to default measure (DD) as in equation (v).
t
trD
V
NP
A
AA
t
2ln
2
)(iv
t
trD
V
DD
A
AA
t
2
ln2
)(v
Reference
Acharya, V, V., Schaefer, S., and Zhang, Y. (2008), “Liquidity Risk and Correlation Risk: A Clinical
Study of the General Motors and Ford Downgrade of May 2005,”
(www.hbs.edu/units/finance/pdf/Acharya_Schaefer_Zhang.pdf).
Allen,F., and Gale,G. (2006), “Systemic Risk and Regulation,”
(fic.wharton.upenn.edu/fic/papers/05/0524.pdf).
Alexopoulou, L., Bunda, I., and Ferrando, A. (2009), “Determinants of Government Bond Spreads in
New EU Countries,” Working Paper Series, No.1093, European Central Bank.
Ammann, M. (2001), Credit Risk Valuation: Methods, Models, and Applications (Second Edition), Springer.
Ashcraft, A, B., and J, A, C, Santos. (2009), “Has the CDS Market Lowered the Cost of Corporate
Debt?,” Journal of Monetary Economics, Vol.56, No.4, pp.514-523.
Baba, N., and Inada, M. (2009), “Price Discovery of Subordinated Credit Spreads for Japanese Mega-
banks: Evidence from Bond and Credit Default Swap Markets,” Journal of International Financial Markets, Institutions & Money, Vol.19, pp.616-632.
Barath, T, S. and Shumway, T. (2008), “Forecasting Default with the Merton Distance to Default
Model,” The Review of Financial Studies, Vol.21, Issue.3, pp.1339-1369.
Baum, F, C., and Wan, C. (2010), “Macroeconomic Uncertainty and Credit Default Swap Spreads,”
Applied Financial Economics, Vol.20, Issue.15, pp.1163-1171.
BIS (2008), “Credit Risk Transfer- Developments form 2005 to 2007,”.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 20 -
Black, F., and Cox, C, J. (1976), “Valuing Corporate Securities: Some Effects of Bond Indenture
Provisions,” The Journal of Finance, Vol.31, No.2, pp.351-367.
Black, F., and Scholes, M. (1972), “The Valuation of Option Contracts and a Test of Market Efficiency,”
The Journal of Finance, Vol.27, pp.399-417.
----------. (1973), “The Pricing of Options and Corporate Liabilities,” Journal of Political Economy,
Vol.81, pp.637-659.
Blanco, R., Brennan, S., and Marsh, I, W. (2005), “An Empirical Analysis of the Dynamic Relationship
between Investment-grade Bonds and Credit Default Swaps,” The Journal of Finance, Vol.60,
No.5, pp.2255-2281.
Bongaerts, D., Jong, D, F., and Driessen, J. (2008), “Liquidity and Liquidity Risk Premia in the CDS
Market,” (www1.fee.uva.nl/pp/bin/895fulltext.pdf).
----------. (2009), “Derivative Pricing with Liquidity Risk: Theory and Evidence from Credit Default Swap Market,” (idei.fr/doc/conf/fbf/driessen.pdf).
Brunnermeier,M,K. (2008), “Deciphering the Liquidity and Credit Crunch 2007-08,” Working paper.
Bystrӧm, H. (2006), “CreditGrades and the iTraxx CDS Index Market,” Financial Analysts Journal, Vo.62, No.6, pp.65-76.
Chau-Lau, A, J. (2006), Fundamentals-Based Estimation of Default Probabilities: A Survey,” IMF Working Paper, WP/06/149, International Monetary Fund.
Collin-Dufresne, P., and Goldstein, S, P. (2001), “Do Credit Spreads Reflect Stationary Leverage
Ratios?,” The Journal of Finance, Vol.56, No.5, pp.1926-1957.
Collin-Dufresne, P., Goldstein, S, P., and Martin, S, J. (2001), “The Determinants of Credit Spread
Changes,” The Journal of Finance, Vol.56, No.6, pp.2177-2207.
Das, R, S., Hanouna, P., and Sarin, A. (2009), “Accounting-based versus Market-based Cross-sectional
Models of CDS Spreads,” Journal of Banking & Finance, Vol.33, Issue.4, pp.719-730.
Di Cesare, A., and Guazzarotti, G. (2010), “An Analysis of the Determinants of Credit Default Swap
Spread Changes Before and During the Subprime Financial Turmoil,” Working Paper, No.749,
Banca D’Italia.
Driessen, J. (2005), “Is Default Event Risk Priced in Corporate Bonds?,” The Review of Financial
Studies, Vol.18, Issue.1, pp.165-195.
Du, Y. and Suo, W. (2007), “Assessing Credit Quality from the Equity Market: Can a Structural
Approach Forecast Credit Ratings?,” Canadian Journal of Administrative Science, Vol.24, pp.212-
228.
Duffee, G. (1998), “The Relationship between Treasury Yields and Corporate Bond Yield Spreads,”
The Journal of Finance, Vol.53, Vol.6, pp.2225-2241.
Edwards, S. (1984), “LDC Foreign Borrowing and Default Risk: An Empirical Investigation,1976-80,”
The American Economic Review, Vol.74, No.4, pp.726-734.
Elton, E, J., Gruber, M, J., Agrawal, D., and Mann, C. (2001), “Explaining the Rate Spread on
Corporate Bonds,” The Journal of Finance, Vol.56, No.1, pp.247-277.
Ericsson, J., Jacobs, K., and Oviedo, R. (2009), “The Determinants of Credit Default Swap Premia,”
Journal of Financial and Quantitative Analysis, Vol.44, No.1, pp.109-132.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 21 -
Ferrucci, G. (2003), “Empirical Determinants of Emerging Market Economies’ Sovereign Bond
Spreads,” Working Paper Series, No.5, Bank of England.
Forte, S., and Lovreta, L. (2009), “Credit Risk Discovery in the Stock and CDS Markets: Who Leads,
When, and Why?,”
(http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1183202).
Forte, S., and Peña, I, J. (2009), “Credit Spreads: An Empirical Analysis on the Information Content of
Stocks, Bonds, and CDS,” Journal of Banking & Finance, Vol.33, Issue.11, pp.2013-2025.
Gai, P., Cameron, G., and Tan, Y, K. (2009), “Sovereign Risk in the Classical Gold Standard Era,” The
Economic Record, Vol.85, No.271, pp.401-416.
Greatrex, A, C. (2008), “The Credit Default Swap Market’s Determinants,” Discussion Paper Series No-2008-05, Fordham University Department of Economics.
Gropp, R., Vesala, J., and Vulpes, G. (2002), “Equity and Bond Market Signals as Leading Indicators of
Bank Fragility,” Working Paper Series, No.150, European Central Bank.
Hakenes,H., and Schnabel,I. (2008), “Credit Risk Transfer in Banking Markets with Hard and Soft
Information,” Joint Deutsche Bundesbank-CEPR-CFS conference Frankfurt am Main, 11-12 December 2008, Deutsche Bundesbank.
----------. (2009), “The Regulation of Credit Derivative Market,” in Macroeconomic Stability and Financial Regulation: Key Issues for the G20, Edit by Dewatripont.M., Freixas.X., and Portes.R.
Hellwig, M. (1994), “Liquidity Provision, Banking and the Allocation of Interest Rate Risk,” European Economic Review, Vol.38, pp.1363-1389.
----------. (2008), “Systemic Risk in the Financial Sector: An Analysis of the Subprime-Mortgage
Financial Crisis,” Preprints of the Max Planck Institute for Research on Collective Goods.
IMF. (2009), “Global Financial Stability Report,” April 2009.
Inaba,K. (2007), “3mega-kou no kurejitto supureddo no kettei youin,” Bank of Japan Working Paper Series, No.07-J-10 (in Japanese).
IOSCO. (2009), “Unregulated Financial Markets and Products- Final Report,”.
Ito, T., and Harada, K. (2004), “Credit Derivatives Premium as a New Japan Premium,” Journal of Money, Credit and Banking, Vol.36, No.5, pp.965-968.
Longstaff, A, F., Mithal, S., and Neis, E. (2005). “Corporate Yield Spreads: Default Risk or Liquidity?
New Evidence from the Credit Default Swap Market,” Journal of Finance, Vol.60, pp.2213-2253.
Longstaff, A, F., and Schwartz, S, E. (1995), “A Simple Approach to Valuing Risky Fixed and Floating
Rate Debt,” The Journal of Finance, Vol. 50, No. 3, pp.789-819.
Martinez-Zarzoso, I., and Bengochea-Morancho, A. (2004), “Pooled Mean Group Estimation of an
Environmental Kuznets Curve for CO2,” Economics Letters, Vol.82, pp.121-126.
Merton, R, C. (1974), “On the Pricing of Corporate Debt: The Risk Structure of Interest Rates,” The Journal of Finance, Vol.29, pp.449-470.
Nakashima, K., and Saito, M. (2009), “Credit Spreads on Corporate Bonds and the Macroeconomy in
Japan,” Journal of The Japanese and International Economies, Vol.23, pp.309-331.
Nashikkar, A., Subrahmanyam, G, M., and Mahonti, S. (2009), “Liquidity and Arbitrage in the Market
for Credit Risk,” (pages.stern.nyu.edu/~msubrahm/papers/CDSPaper.pdf).
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 22 -
Ooyama, S., and T. Sugimoto. (2007), “Nihon ni Okeru Kurejitto Supureddo no Hendou Youin,” Bank of Japan Working Paper Series, No.07-J-1 (in Japanese).
Ooyama, S., and Y. Hongo. (2010), “Nihon no Shasai Hakko Supureddo no Hendo Yoin,” Bank of Japane Working Paper Series, No.10-J-10 (in Japanese).
Ötker-Robe, İnci, and Podpiera, J. (2010), “The Fundamental Determinants of Credit Default Risk for
European Large Complex Financial Institutions,” IMF Working Paper, WP/10/153, International
Monetary Fund.
Pan, J., and Singleton, J, K. (2006), Interpreting Recent Changes in the Credit Spreads of Japanese
Banks,” IMES Discussion Paper Series, No.2006-E-21, Bank of Japan.
Partnoy, F., and Skeel, Jr, A, D. (2007), “The Promise and Perils of Credit Derivatives,” Legal Studies Research Paper No.07-74, University of San Diego School of Law.
Pesaran, M, H., and Shin, Y. (1998), “An Autoregressive Distributed-Lag Modelling Approach to
Cointegration Analysis,” Econometrics and Economic Theory in the 20th
Century: The Ragner Frisch Centennial Symposium, Cambridge University Press.
Pesasan, M, H,. Shin, Y,. and Smith, R, P. (1999), ”Pooled Mean Group Estimation of Dynamic
Heterogeneous Panels,” Journal of the American Statistical Association, Vol.94, pp.621-634.
Pesaran, M, H., and Smith, P, R. (1995), “Estimating Long-run Relationships From Dynamic
Heterogeneous Panels,” Journal of Econometrics, Vol.68, pp.79-113.
Pynnӧnen, S., Hogan, W., and Batten, J. (2004), “Dynamic Equilibrium Correction Modelling of Credit
Spreads. The Case of Yen Eurobonds,” (lipas.uwasa.fi/~sjp/articles/phb_acta_wasaensia_122_187-
204.pdf).
Shirasu, Y. and Y. Yonezawa. (2007), “Shasai Ryuutuu Sizyo ni Okeru Shasai Supureddo Hendo Yoin
no Jissho Bunseki,” Discussion Paper, No.2007-2, Financial Research Center, Financial Services
Agency, Japan (in Japanese).
Simons, D., and Rolwes, F. (2009), “Macroeconomic Default Modelling and Stress Testing,”
International Journal of Central Banking, Vol.5, No.3, pp.177-204.
Sommar, Å, P., and Shahnazarian, H. (2009), “Interdependenicies between Expected Default Frequecny
and the Macro Economy,” International Journal of Central Banking, Vol.5, No.3, pp.83-110.
Tang, Y, D., and Yan, H. (2008), “Liquidity and Credit Default Swap Spreads,”
(http://papers.ssrn.com/sol3/papers.cfm?abstract_id=891263&rec=1&srcabs=675641).
Vassalou, M., and Xing, Y. (2004), “Default Risk in Equity Returns,” The Journal of Finance, Vol.59,
No.2, pp.831-868.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 23 -
Table 1: Definition of Variables
Nam
e o
f v
ari
ab
leD
efi
nit
ion
So
urc
e
CD
SS
Lev
el v
alu
e o
f C
DS
sp
read
J-C
DS
data
base p
ub
lish
ed
by
To
ky
o F
inan
cia
l E
xch
an
ge I
nc.
ln(C
DS
S)
Natu
ral lo
gari
thm
of
CD
SS
〃
⊿C
DS
SC
DS
S a
t ti
me t
min
us C
DS
S a
t ti
me t
-1〃
⊿ln
(CD
SS
)ln
(CD
SS
) a
t ti
me t
min
us l
n(C
DS
S)
at
tim
e t
-1〃
LE
VE
RA
GE
{D
eb
t T
ota
l ÷
(Deb
t T
ota
l+M
ark
et
Valu
e o
f E
qu
ity
)} is t
ran
sfo
med
to
mo
nth
ly d
ata
usin
g lin
ear
inte
rpo
lati
on
meth
od
. F
inan
cia
l d
ata
are
sem
ian
nu
al.
Co
llecte
d b
y A
str
a M
an
ag
er
data
base (
QU
ICK
)
VO
LA
TIL
ITY
Calc
ula
ted
based
on
th
e p
roced
ure
in
Ap
pen
dix
.〃
RF
Inte
rest
rate
on
10 y
ear
go
vern
men
t b
on
d in
th
e s
eco
nd
ary
mark
et
〃
TS
Inte
rest
rate
on
10 y
ear
go
vern
men
t b
on
d -
in
tere
st
rate
on
1 y
ear
go
vern
men
t b
on
d〃
TO
PIX
Dif
fere
ce o
f n
atu
ral lo
gari
thm
of
TO
PIX
In
dex
at
tim
e t
co
mp
are
d t
o p
rev
iou
s d
ay
〃
ln(M
V)
Natu
ral lo
gari
thm
of
Mark
et
Valu
e o
f E
qu
ity
〃
SW
AP
TIO
NIm
plied
vo
lati
lity
of
Sw
ap
tio
n(
5 y
ears
, 1 m
on
th m
atu
rity
)B
loo
mb
erg
VIX
Imp
lied
vo
lati
tily
of
S&
P500 in
dex
〃
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 24 -
Table 2: List of Sample
(1)Name of reference entity
Security ID Name of firm Security ID Name of firm
4005 Sumitomo Chemical Company, Limited 8031 Mitsui & Co., Ltd.
4183 Mitsui Chemicals, Inc. 8053 Sumitomo Corporation
5001 Nippon Oil Corporation 8058 Mitsubishi Corporation
5401 Nippon Steel Corporation 8515 Aiful Corporation
5802 Sumitomo Electric Industries, Ltd 8564 Takefuji Corporation
6501 Hitachi, Ltd. 8572 Acom Co., Ltd.
6502 Toshiba Corporation 8574 Promise Co., Ltd.
6503 Mitsubishi Electric Corporation 8591 ORIX Corporation
6701 NEC Corporation 8801 Mitsui Fudosan Co., Ltd.
6702 Fujitsu Limited 8802 Mitsubishi Estate Co., Ltd.
6752 Panasonic Corporation 9005 Tokyu Corporation
6753 Sharp Corporation 9041 Kintetsu Corporation
6758 Sony Corporation 9042 Hankyu Hanshin Holdings, Inc.
6764 SANYO Electric Co., Ltd. 9202 All Nippon Airways Co., Ltd.
6952 Casio Computer Co., Ltd. 9205 Japan Airlines Corporation
7011 Mitsubishi Heavy Industries, Ltd. 9433 KDDI Corporation
7012 Kawasaki Heavy Industries, Ltd. 9437 NTT DOCOMO, INC.
7201 Nissan Motor Co.,Ltd. 9501 The Tokyo Electric Power Company,
7203 Toyota Motor Corporation 9502 Chubu Electric Power Co.,Inc.
7267 Honda Motor Co., Ltd. 9503 The Kansai Electric Power Company,
7269 SuzukiI Motor Corporation 9531 Tokyo Gas Co., Ltd.
7731 Nikon Corporation 9532 Osaka Gas Co., Ltd.
7752 Ricoh Company, Ltd.
(2)By rating class
# of firms share # of firms share # of firms share
AAA 1 2.2% 1 2.2% 1 2.2%
AA+ 7 15.6% 7 15.6% 7 15.6%
AA 4 8.9% 4 8.9% 3 6.7%
AA- 6 13.3% 7 15.6% 8 17.8%
A+ 6 13.3% 8 17.8% 7 15.6%
A 8 17.8% 8 17.8% 8 17.8%
A- 5 11.1% 3 6.7% 4 8.9%
BBB+ 4 8.9% 2 4.4% 1 2.2%
BBB 3 6.7% 4 8.9% 3 6.7%
BB+ 1 2.2% 1 2.2% 1 2.2%
BB- 0 0.0% 0 0.0% 1 2.2%
CCC+ 0 0.0% 0 0.0% 1 2.2%
Total 45 100.0% 45 100.0% 45 100.0%
Rating is the R&I's long-term domestic-currency credit rating.
2004/4/1 2007/6/29 2009/9/30Rating
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 25 -
Table3: Summary Statistics
(1)
Basic
Sta
tisti
cs
Nam
e o
f v
ari
ab
leO
bs
Mean
S.D
.M
inM
ax
CD
SS
60,7
50
89.2
56
334.0
47
2.5
00
7654.9
60
ln(C
DS
S)
60,7
50
3.3
56
1.2
04
0.9
16
8.9
43
LE
VE
RA
GE
60,7
50
0.1
57
0.0
94
0.0
18
0.9
04
VO
LA
TIL
ITY
60,7
50
0.2
89
0.1
44
0.0
57
1.3
57
RF
60,7
50
0.0
15
0.0
02
0.0
12
0.0
20
SW
AP
TIO
N60,6
15
46.8
35
10.4
04
25.3
00
84.0
00
TS
60,7
50
0.0
12
0.0
02
0.0
07
0.0
19
TO
PIX
60,7
50
0.0
00
0.0
16
-0.1
00
0.1
29
ln(M
V)
60,7
50
27.9
96
0.9
42
23.3
06
31.0
36
VIX
58,7
70
20.4
74
12.0
24
9.8
90
80.8
60
(2)
Co
rrela
tio
n
CD
SS
ln(C
DS
S)
Levera
ge
ζA
rfS
wa
pti
on
TS
⊿ln
(TO
PIX
)L
n(M
V)
VIX
CD
SS
1
ln(C
DS
S)
0.5
914*
1
LE
VE
RA
GE
0.5
606*
0.6
170*
1
VO
LA
TIL
ITY
0.1
881*
0.4
829*
0.1
039*
1
RF
-0.1
661*
-0.3
275*
-0.2
311*
-0.1
136*
1
SW
AP
TIO
N0.0
193*
0.1
206*
0.1
344*
-0.0
092*
-0.3
888*
1
TS
-0.1
329*
-0.3
194*
-0.0
951*
-0.3
995*
0.2
970*
0.3
429*
1
TO
PIX
0.0
01
-0.0
138*
-0.0
184*
-0.0
182*
0.0
219*
-0.0
05
0.0
774*
1
ln(M
V)
-0.3
861*
-0.5
299*
-0.4
837*
-0.1
422*
0.1
793*
-0.1
343*
0.0
220*
0.0
107*
1
VIX
0.2
933*
0.5
918*
0.2
971*
0.6
419*
-0.3
397*
0.1
747*
-0.5
276*
-0.0
979*
-0.1
856*
1
* r
efe
rs t
o t
he s
tati
sti
cal sig
nif
ican
ce a
t th
e 5
% lev
el.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 26 -
Table3 (continued)
(3)
Un
it R
oo
t T
est
(3)
-1. F
irm
sp
ecif
ic v
ari
ab
les
1%
5%
10%
1%
5%
10%
CD
SS
63
045
00
ln(C
DS
S)
00
043
20
VO
LA
TIL
ITY
73
045
00
LE
VE
RA
GE
00
037
30
ln(M
V)
11
038
40
1%
5%
10%
1%
5%
10%
CD
SS
10
045
00
ln(C
DS
S)
00
045
00
VO
LA
TIL
ITY
41
045
00
LE
VE
RA
GE
11
045
00
ln(M
V)
11
045
00
Th
e n
um
ber
in t
he t
ab
le is t
he n
um
ber
of
firm
s f
or
wh
ich
th
e n
ull h
yp
ote
sis
is r
eje
cte
d. T
ota
l n
um
ber
of
sam
ple
s is 4
5.
(3)
-2. v
ari
ab
les c
om
mo
n t
o f
irm
s
RF
-2.4
7-9
.48
**
*-2
.78
-38.8
3*
**
SW
AP
TIO
N-2
.23
-0.4
1-4
.31
**
*-5
0.8
9*
**
TS
-2.7
7*
-8.6
3*
**
-2.8
7-3
9.0
2*
**
TO
PIX
-36.9
2*
**
-0.5
2-3
7.9
7*
**
-358.7
9*
**
VIX
-1.7
3-2
7.8
5*
**
-2.7
7-4
4.1
7*
**
**
*, *
*, an
d *
refe
r to
th
e s
tati
sti
cal sig
nif
ican
ce a
t 1%
, 5%
, an
d 1
0%
resp
ecti
vely
.
DF
-GL
S t
est
PP
test
Lev
el
Fir
st
dif
fere
nce
Lev
el
Fir
st
dif
fere
nce
Lev
el
Fir
st
dif
fere
nce
DF
-GL
S t
est
PP
test
Lev
el
Fir
st
dif
fere
nce
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 27 -
Table4: Time Series Regression (Level Variable)
ad
j-R
2, F
-valu
e, o
bs, co
ef,
s.e
. are
cro
ss-s
ecti
on
al av
era
ge o
f esti
mati
on
resu
lts f
or
each
fir
m. F
igu
res in
pare
nth
eses a
re c
ross-s
ecti
on
la m
ed
ian
fo
r co
ef
an
d s
.e. R
ob
ust
sta
nd
ard
err
os a
re u
sed
in
OL
S. t-
valu
e a
nd
p-v
alu
e a
re t
he t
-valu
e
an
d p
-valu
e p
ert
ain
ing
to
th
e c
on
sta
nt
term
fro
m t
he O
LS
esti
mati
on
in
wh
ich
dep
en
den
t v
ari
ab
le c
on
sis
t o
f th
e p
ara
mete
r esti
mate
s d
eri
ved
fro
m O
LS
reg
ressio
ns f
or
each
fir
m w
hile e
xpla
nato
ry v
ari
ab
le is a
co
nsta
nt
term
. T
he c
oef
an
d
s.e
. fo
r S
QR
(RF
) are
exp
ressed
dev
ided
by
10000.
t-v
alu
ep
-valu
et-
valu
ep
-valu
et-
valu
ep
-valu
e
LE
VE
RA
GE
859
<524>
59
<40>
3.4
10
0.0
01
1,1
59
<795>
232
<136>
1.2
50
0.2
17
1,1
04
<794>
226
<135>
1.9
00
0.0
65
VO
LA
TIL
ITY
218
<131>
22
<6.5
6>
3.0
20
0.0
04
174
<89>
20
<6.6
4>
3.5
80
0.0
01
62
<48>
23
<11>
2.5
70
0.0
14
RF
-8,9
82
<-2
,479>
955
<364>
-2.7
30
0.0
09
-100,9
80
<-3
3,2
44>
11,4
75
<5,1
46>
-3.9
60
0.0
00
-92,3
53
<-3
8,5
52>
10,8
31
<5,1
13>
-4.1
60
0.0
00
SQ
R(R
F)
--
279
<100>
36
<16>
4.0
10
0.0
00
249
<107>
33
<16>
4.3
00
0.0
00
TS
--
13,2
35
<935>
1,2
42
<489>
1.6
90
0.0
98
16,0
06
<2,4
22>
1,2
32
<522>
2.1
80
0.0
35
TO
PIX
--
206
<116>
112
<56>
5.7
90
0.0
00
254
<133>
119
<58>
5.6
90
0.0
00
Ln
(MV
)-
-3.9
0<
-7.2
0>
35
<14>
0.0
30
0.9
80
27
<25>
33
<16>
0.2
90
0.7
72
SW
AP
TIO
N-
--0
.502
<-0
.053>
0.2
07
<0.0
92>
-0.8
60
0.3
96
-1.0
09
<-0
.278>
0.2
14
<0.0
96>
-2.1
00
0.0
41
VIX
--
--
1.6
83
<0.4
91>
0.3
67
<0.1
81>
2.1
80
0.0
34
co
nst
an
t24.2
7<
-23.1
9>
20.0
0<
7.8
2>
0.3
60
0.7
19
411
<107>
996
<371>
0.0
90
0.9
27
-306
<-3
31>
930
<483>
-0.1
20
0.9
08
ad
j-R
2
F-v
alu
e
ob
s
t-v
alu
ep
-valu
et-
valu
ep
-valu
et-
valu
ep
-valu
e
LE
VE
RA
GE
14
<9.1
88>
0.5
64
<0.4
41>
4.5
70
0.0
00
6.2
93
<1.2
04>
1.8
74
<1.5
63>
0.9
70
0.3
38
5.0
20
<5.6
08>
1.9
27
<1.5
72>
0.8
50
0.4
01
VO
LA
TIL
ITY
3.4
32
<3.3
64>
0.1
29
<0.1
19>
13.9
70
0.0
00
2.5
07
<2.2
51>
0.1
30
<0.1
24>
13.7
90
0.0
00
1.8
62
<1.7
57>
0.1
67
<0.1
52>
12.2
10
0.0
00
RIS
KF
RE
E-6
2<
-57>
6.6
71
<6.3
83>
-7.5
90
0.0
00
-613
<-5
84>
83.0
13
<77>
-11.0
00
0.0
00
-603
<-6
31>
80
<77>
-12.4
20
0.0
00
SQ
R(R
F)
--
1.8
87
<1.7
40>
0.2
57
<0.2
42>
11.6
90
0.0
00
1.8
40
<1.8
36>
0.2
49
<0.2
39>
13.2
60
0.0
00
TS
--
-33
<-2
7>
7.8
42
<7.1
76>
-2.4
00
0.0
21
-16
<-5
.017>
8.3
92
<8.0
06>
-1.2
80
0.2
07
TO
PIX
--
1.3
30
<1.3
10>
0.7
08
<0.6
89>
11.5
50
0.0
00
1.6
11
<1.6
20>
0.7
58
<0.7
50>
13.1
10
0.0
00
ln(M
V)
--
-0.9
61
<-1
.499>
0.2
20
<0.1
83>
-1.3
20
0.1
93
-0.9
50
<-1
.047>
0.2
19
<0.1
86>
-1.5
60
0.1
26
SW
AP
TIO
N-
-0.0
06
<0.0
05>
0.0
01
<0.0
01>
4.5
00
0.0
00
0.0
04
<0.0
04>
0.0
02
<0.0
01>
3.5
70
0.0
01
VIX
--
--
0.0
09
<0.0
07>
0.0
02
<0.0
02>
3.4
80
0.0
01
co
nst
an
t1.8
68
<1.4
24>
0.1
47
<0.1
35>
6.1
60
0.0
00
34
<51>
6.3
21
<5.2
02>
1.6
30
0.1
11
34
<32>
6.3
14
<5.3
42>
1.9
10
0.0
63
ad
j-R
2
F-v
alu
e
ob
s1,3
50
1,3
47
1,3
03
-
0.7
40.8
30.8
5
1,7
99
1,0
74
1,0
39
-
--
--
--
--
--
--
1,3
50
1,3
47
1,3
03
y=
ln(C
DS
S)
co
ef
s.e
.co
ef
s.e
.co
ef
s.e
.
545
356
343
--
--
--
--
0.6
70.7
70.7
9
--
--
--
y=
CD
SS
co
ef
s.e
.co
ef
s.e
.co
ef
s.e
.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 28 -
Table 5: Time Series Regression (Differenced Variable)
ad
j-R
2, F
-valu
e, o
bs, co
ef,
s.e
. are
cro
ss-s
ecti
on
al av
era
ge o
f esti
mati
on
resu
lts f
or
each
fir
m. F
igu
res in
pare
nth
eses a
re c
ross-s
ecti
on
la m
ed
ian
fo
r co
ef
an
d s
.e. R
ob
ust
sta
nd
ard
err
os a
re u
sed
in
OL
S. t-
valu
e a
nd
p-v
alu
e a
re t
he t
-valu
e
an
d p
-valu
e p
ert
ain
ing
to
th
e c
on
sta
nt
term
fro
m t
he O
LS
esti
mati
on
in
wh
ich
dep
en
den
t v
ari
ab
le c
on
sis
t o
f th
e p
ara
mete
r esti
mate
s d
eri
ved
fro
m O
LS
reg
ressio
ns f
or
each
fir
m w
hile e
xpla
nato
ry v
ari
ab
le is a
co
nsta
nt
term
. T
he c
oef
an
d
s.e
. fo
r S
QR
(RF
) are
exp
ressed
as d
ivid
ed
by
10000.
t-v
alu
ep
-valu
et-
valu
ep
-valu
et-
valu
ep
-valu
e
⊿L
EV
ER
AG
E213
<127>
76
<39>
4.9
00
0.0
00
130
<38>
90
<53>
2.7
20
0.0
09
131
<25>
94
<46>
2.6
40
0.0
11
⊿V
OL
AT
ILIT
Y28
<11>
27
<13>
2.7
50
0.0
09
24
<8.4
80>
27
<13>
2.4
20
0.0
20
26
<7.9
94>
30
<13>
2.4
70
0.0
17
⊿R
F-3
30
<-2
35>
447
<187>
-2.4
50
0.0
18
-8.3
33
<25>
469
<199>
-0.0
70
0.9
44
-80
<-5
.953>
470
<186>
-0.7
90
0.4
35
⊿S
QR
(RF
)-
--9
.627
<-2
.826>
73
<32>
-0.4
30
0.6
66
-25
<-7
.387>
75
<31>
-0.9
70
0.3
37
TS
--
-45
<-5
0>
100
<28>
-1.0
10
0.3
17
-29
<-3
9>
104
<28>
-0.5
80
0.5
67
TO
PIX
--
-21
<-2
1>
21
<9.1
12>
-2.4
40
0.0
19
-5.9
15
<-1
3>
23
<9.6
36>
-0.5
40
0.5
92
ln(M
V)
--
-1.3
22
<-0
.287>
1.0
32
<0.2
81>
-2.6
70
0.0
11
-1.2
94
<-0
.187>
1.0
58
<0.2
65>
-2.4
30
0.0
19
⊿S
WA
PT
ION
--
0.0
13
<0.0
03>
0.0
62
<0.0
22>
0.8
50
0.4
00
0.0
05
<0.0
01>
0.0
67
<0.0
23>
0.2
60
0.7
97
⊿V
IX-
--
-0.2
12
<0.0
74>
0.1
33
<0.0
56>
3.6
50
0.0
01
co
nst
an
t0.2
58
<0.0
21>
0.1
56
<0.0
68>
1.9
90
0.0
53
37
<8.4
55>
28
<8.3
60>
2.7
40
0.0
09
36
<5.9
62>
29
<7.9
05>
2.4
80
0.0
17
ad
j-R
2
F-v
alu
e
ob
s
t-v
alu
ep
-valu
et-
valu
ep
-valu
et-
valu
ep
-valu
e
⊿L
EV
ER
AG
E1.8
76
<1.6
56>
0.4
51
<0.3
61>
8.2
20
0.0
00
0.7
16
<0.4
26>
0.5
20
<0.4
24>
3.5
00
0.0
01
0.6
72
<0.3
54>
0.5
18
<0.4
16>
3.7
40
0.0
01
⊿V
OL
AT
ILIT
Y0.3
23
<0.3
05>
0.1
61
<0.1
56>
11.6
50
0.0
00
0.2
71
<0.2
60>
0.1
55
<0.1
49>
10.0
00
0.0
00
0.2
42
<0.2
29>
0.1
52
<0.1
44>
9.1
70
0.0
00
⊿R
F-6
.755
<-7
.354>
3.0
55
<2.8
96>
-12.7
20
0.0
00
-2.0
64
<-1
.608>
3.0
18
<2.9
31>
-5.4
70
0.0
00
-2.7
03
<-2
.523>
3.0
52
<2.9
94>
-7.3
70
0.0
00
⊿S
QR
(RF
)-
-0.3
27
<0.3
74>
0.6
93
<0.6
56>
4.1
20
0.0
00
0.2
04
<0.2
09>
0.7
03
<0.6
72>
2.5
50
0.0
14
TS
--
-0.8
73
<-0
.870>
0.3
88
<0.3
44>
-15.4
10
0.0
00
-0.7
64
<-0
.739>
0.3
76
<0.3
41>
-14
0.0
00
TO
PIX
--
-0.3
08
<-0
.319>
0.0
92
<0.0
92>
-16.3
90
0.0
00
-0.1
73
<-0
.185>
0.0
96
<0.0
93>
-9.4
20
0.0
00
ln(M
V)
--
-0.0
02
<-0
.002>
0.0
03
<0.0
03>
-2.6
20
0.0
12
-0.0
02
<-0
.002>
0.0
03
<0.0
03>
-2.5
70
0.0
14
⊿S
WA
PT
ION
--
0.0
003
<0.0
00>
0.0
00
<0.0
00>
8.5
60
0.0
00
0.0
002
<0.0
00>
0.0
00
<0.0
00>
5.7
10
0.0
00
⊿V
IX-
--
-0.0
01
<0.0
01>
0.0
01
<0.0
01>
24
0.0
00
co
nst
an
t0.0
01
<0.0
01>
0.0
01
<0.0
01>
9.2
90
0.0
00
0.0
71
<0.0
62>
0.0
93
<0.0
83>
3.0
80
0.0
04
0.0
70
<0.0
55>
0.0
95
<0.0
85>
2.9
40
0.0
05
ad
j-R
2
F-v
alu
e
ob
s1,3
50
1,3
45
1,2
57
--
0.0
40.0
60.0
7
10.3
36.3
96.3
4
--
--
--
--
--
--
1,3
50
1,3
45
1,2
57
y=⊿
ln(C
DS
S)
co
ef
s.e
.co
ef
s.e
.co
ef
s.e
.
7.4
95.8
05.0
0
--
--
--
--
0.0
30.0
60.0
6
--
--
--
y=⊿
CD
SS
co
ef
s.e
.co
ef
s.e
.co
ef
s.e
.
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 29 -
Table6: PMG Model (without Restriction, Full Sample)
χ2 is t
he t
est
sta
tisti
cs f
or
the n
ull h
yp
oth
esis
th
at
para
mete
rs a
re h
om
og
en
eo
us, an
d f
igu
res in
pare
nth
eses a
re a
sso
cia
ted
p-v
alu
e. T
he c
oef
an
d s
.e. fo
r S
QR
(RF
) an
d ⊿
SQ
R(R
F)
are
exp
ressed
as d
ivid
ed
by
10000. F
igu
res in
lon
g-r
un
eq
uilib
riu
m e
qu
ati
on
are
resu
lts a
bo
ut
ho
mo
gen
eo
us p
ara
mete
rs. c
oef
of
oth
er
exp
lan
ato
ry v
ari
ab
les a
re c
ross-s
ecti
on
al av
era
ge o
f co
eff
icie
nts
fo
r each
fir
ms. s
.e., z
-valu
e, an
d p
-valu
e o
f th
ese v
ari
ab
les a
re t
he
sta
nd
ard
err
or,
z-v
alu
es a
nd
p-v
alu
es r
ela
tin
g t
o t
he c
on
sta
nt
term
in
th
e O
LS
esti
mati
on
in
wh
ich
th
e p
ara
mete
r esti
mate
s o
f each
issu
e a
re t
he d
ep
en
den
t v
ari
ab
le w
hile t
he c
on
sta
nt
term
is e
xpla
nato
ry v
ari
ab
le. ad
j-R
2 is
cro
ss-s
ecti
on
al av
era
ge. T
he r
esu
lts o
n c
on
sta
nt
term
s a
re o
mit
ted
.
Pan
el A
: d
ep
en
den
t v
ari
ab
le=
⊿C
DS
S
co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)
lon
g-t
erm
eq
uati
on
LE
VE
RA
GE
491
33
14.7
50
0.0
00
920
105
8.7
20
0.0
00
629
57
11.0
40
0.0
00
VO
LA
TIL
ITY
289
10
27.9
50
0.0
00
205
16
12.9
70
0.0
00
177
8.7
45
20.2
00
0.0
00
RF
-2,7
03
616
-4.3
90
0.0
00
-196,0
53
15,7
09
-12.4
80
0.0
00
-106,4
00
8,1
33
-13.0
80
0.0
00
SQ
R(R
F)
(ii)
--
--
657
51
12.9
20
0.0
00
355
26
13.6
60
0.0
00
TS
--
--
-20,2
36
1,3
84
-14.6
20
0.0
00
-10,1
69
656
-15.5
00
0.0
00
TO
PIX
--
--
-7,8
58
455
-17.2
70
0.0
00
-3,7
19
180
-20.6
30
0.0
00
ln(M
V)
--
--
69
15
4.7
00
0.0
00
41
7.6
10
5.4
40
0.0
00
SW
AP
TIO
N-
--
-5.3
94
0.3
67
14.7
00
0.0
00
2.9
99
0.1
72
17.4
10
0.0
00
sh
ort
-term
eq
uati
on
ad
just
men
t p
ara
mer
ter
(ϕi)
-0.0
06
0.0
01
-7.4
50
0.0
00
487.5
5(0
.0000)
-0.0
05
0.0
01
-8.1
30
0.0
00
474.9
1(0
.0000)
-0.0
07
0.0
01
-8.5
70
0.0
00
805.8
0(0
.0000)
⊿L
EV
ER
AG
E205
43
4.7
20
0.0
00
646.9
5(0
.0000)
252
91
2.7
80
0.0
05
282.2
8(0
.0000)
238
97
2.4
40
0.0
15
279.8
5(0
.0000)
⊿V
OL
AT
ILIT
Y24
10
2.3
90
0.0
17
157.0
5(0
.0000)
20
11
1.8
30
0.0
67
164.9
9(0
.0000)
22.3
91
12
1.8
70
0.0
62
174.0
4(0
.0000)
⊿R
F-3
25
136
-2.3
90
0.0
17
43.2
8(0
.5023)
511
275
1.8
60
0.0
63
11.2
4(1
.0000)
525
297
1.7
70
0.0
77
12.9
6(1
.0000)
⊿S
QR
(RF
)(i
i)-
--
--
-42
18
-2.3
50
0.0
19
20.2
6(0
.9992)
-59
18.6
97
-3.1
40
0.0
02
21.7
5(0
.9980)
⊿T
S-
--
--
-1,0
34
382
-2.7
10
0.0
07
13.0
6(1
.0000)
-1,0
32
359
-2.8
80
0.0
04
12.2
8(1
.0000)
⊿T
OP
IX-
--
--
23
3.1
10
7.2
80
0.0
00
388.3
5(0
.0000)
23
3.9
61
5.8
20
0.0
00
377.6
0(0
.0000)
⊿ln
(MV
)-
--
--
15
20.7
72
0.7
30
0.4
65
190.1
4(0
.0000)
12
23.1
17
0.5
30
0.5
95
186.4
8(0
.0000)
⊿S
WA
PT
ION
--
--
-0.0
11
0.0
17
0.6
70
0.5
02
41.0
5(0
.5987)
0.0
03
0.0
16
0.1
80
0.8
56
43.2
7(0
.5029)
⊿V
IX-
--
--
--
--
-0.1
94
0.0
56
3.4
50
0.0
01
187.0
3(0
.0000)
log
lik
elih
oo
d
ad
j-R
2
ob
s60,7
49
60,5
24
56,5
64
-146,9
88
-144,9
97
-132,1
44
0.0
40.0
80.2
3
PM
GP
MG
PM
G
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 30 -
Table6 (continued)
Pan
el B
: d
ep
en
den
t v
ari
ab
le=
⊿ln
(CD
SS
)
co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)1
co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)1
co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)1
lon
g-t
erm
eq
uati
on
LE
VE
RA
GE
8.6
04
0.6
58
13.0
80
0.0
00
6.7
16
1.0
23
6.5
70
0.0
00
7.7
84
1.0
13
7.6
80
0.0
00
VO
LA
TIL
ITY
5.9
88
0.2
45
24.4
20
0.0
00
2.7
50
0.1
86
14.8
20
0.0
00
2.9
70
0.1
84
16.1
30
0.0
00
RF
-88
14
-6.1
20
0.0
00
-2,8
97
182
-15.9
00
0.0
00
-2,5
69
175
-14.6
70
0.0
00
SQ
R(R
F)
(ii)
--
--
10
0.5
87
16.2
10
0.0
00
8.5
44
0.5
64
15.1
40
0.0
00
TS
--
--
-314
16
-19.9
30
0.0
00
-281
15
-18.9
50
0.0
00
TO
PIX
--
--
-97
4.7
10
-20.5
50
0.0
00
-73.1
3.9
08
-18.7
10
0.0
00
ln(M
V)
--
--
-0.0
37
0.1
61
-0.2
30
0.8
19
0.0
26
0.1
59
0.1
60
0.8
71
SW
AP
TIO
N-
--
-0.0
54
0.0
03
16.2
90
0.0
00
0.0
54
0.0
03
16.4
00
0.0
00
sh
ort
-term
eq
uati
on
ad
just
men
t p
ara
mer
ter
(ϕi)
-0.0
04
0.0
00
-12.8
10
0.0
00
138.9
2(0
.0000)
-0.0
06
0.0
00
-29.4
00
0.0
00
125.7
4(0
.0000)
-0.0
06
0.0
00
-24.3
50
0.0
00
125.9
9(0
.0000)
⊿L
EV
ER
AG
E1.8
32
0.2
27
8.0
70
0.0
00
368.1
4(0
.0000)
0.8
75
0.3
28
2.6
70
0.0
08
138.2
9(0
.0000)
1.1
87
0.3
60
3.3
00
0.0
01
140.0
9(0
.0000)
⊿V
OL
AT
ILIT
Y0.2
93
0.0
27
10.7
70
0.0
00
212.9
0(0
.0000)
0.1
85
0.0
24
7.7
50
0.0
00
141.7
8(0
.0000)
0.1
57
0.0
23
6.7
70
0.0
00
117.6
5(0
.0000)
⊿R
F-6
.555
0.5
27
-12.4
40
0.0
00
49.3
8(0
.2668)
-7.2
14
0.9
02
-8.0
00
0.0
00
27.7
4(0
.9736)
-7.8
80
1.0
04
-7.8
50
0.0
00
35.3
0(0
.8225)
⊿S
QR
(RF
)(i
i)-
--
--
0.0
49
0.0
76
0.6
50
0.5
17
26.4
9(0
.9830)
-0.1
08
0.0
75
-1.4
50
0.1
46
26.3
7(0
.9838)
⊿T
S-
--
--
4.3
53
0.7
75
5.6
20
0.0
00
21.3
5(0
.9984)
4.4
53
0.8
66
5.1
40
0.0
00
26.2
0(0
.9848)
⊿T
OP
IX-
--
--
0.2
56
0.0
09
27.1
70
0.0
00
93.2
7(0
.0000)
0.2
34
0.0
09
25.5
60
0.0
00
92.2
4(0
.0000)
⊿ln
(MV
)-
--
--
0.0
40
0.0
37
1.0
60
0.2
87
126.1
7(0
.0000)
0.0
74
0.0
37
2.0
10
0.0
45
112.5
4(0
.0000)
⊿S
WA
PT
ION
--
--
-0.0
001
0.0
00
2.4
20
0.0
15
18.1
8(0
.9998)
0.0
0004
0.0
00
0.9
50
0.3
43
22.2
0(0
.9975)
⊿V
IX-
--
--
--
--
-0.0
01
0.0
00
17.1
00
0.0
00
48.3
8(0
.3007)
log
lik
elih
oo
d
ad
j-R
2
ob
s60,7
49
60,5
24
56,5
64
128,4
65
129,7
10
123,4
37
0.0
0-0
.01
0.1
3
PM
GP
MG
PM
G
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 31 -
Table7: PMG Model (With Restriction, Full Sample)
χ2 is t
he t
est
sta
tisti
cs f
or
the n
ull h
yp
oth
esis
th
at
para
mete
rs a
re h
om
og
en
eo
us, an
d f
igu
res in
pare
nth
eses a
re a
sso
cia
ted
p-v
alu
e. T
he c
oef
an
d s
.e. fo
r S
QR
(RF
) an
d ⊿
SQ
R(R
F)
are
exp
ressed
as d
ivid
ed
by
10000. F
igu
res in
lo
ng
-ru
n e
qu
ilib
riu
m e
qu
ati
on
are
resu
lts a
bo
ut
ho
mo
gen
eo
us p
ara
mete
rs. c
oef
of
oth
er
exp
lan
ato
ry v
ari
ab
les a
re c
ross-s
ecti
on
al
av
era
ge o
f co
eff
icie
nts
fo
r each
fir
ms. s
.e., z
-valu
e, an
d p
-valu
e o
f th
ese v
ari
ab
les a
re t
he s
tan
dard
err
or,
z-v
alu
es a
nd
p-v
alu
es r
ela
tin
g t
o t
he c
on
sta
nt
term
in
th
e O
LS
esti
mati
on
in w
hic
h t
he p
ara
mete
r esti
mate
s o
f each
issu
e a
re t
he d
ep
en
den
t v
ari
ab
le w
hile t
he c
on
sta
nt
term
is e
xpla
nato
ry v
ari
ab
le. ad
j-R
2 is c
ross-s
ecti
on
al av
era
ge. T
he r
esu
lts o
n c
on
sta
nt
term
s a
re o
mit
ted
.
Pan
el A
: d
ep
en
den
t v
ari
ab
le=
⊿C
DS
S
co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)
lon
g-t
erm
eq
uati
on
LE
VE
RA
GE
752
85
8.8
90
0.0
00
578
49
11.7
60
0.0
00
VO
LA
TIL
ITY
182
13
14.4
00
0.0
00
162
7.5
53
21.4
20
0.0
00
RF
-172,0
41
12,4
14
-13.8
60
0.0
00
-101,1
62
7,0
67
-14.3
20
0.0
00
SQ
R(R
F)
(ii)
575
40
14.3
30
0.0
00
336
23
14.9
00
0.0
00
TS
-18,1
66
1,1
05
-16.4
40
0.0
00
-9,9
81
577
-17.2
90
0.0
00
TO
PIX
-3,5
78
205
-17.4
30
0.0
00
-1,4
79
94
-15.6
80
0.0
00
ln(M
V)
47
12
4.0
40
0.0
00
30
6.4
02
4.7
50
0.0
00
SW
AP
TIO
N4.4
17
0.2
78
15.9
00
0.0
00
2.6
07
0.1
46
17.8
20
0.0
00
sh
ort
-term
eq
uati
on
ad
just
men
t p
ara
mer
ter
(ϕi)
-0.0
05
0.0
01
-8.5
10
0.0
00
542.6
9(0
.000)
-0.0
08
0.0
01
-8.6
60
0.0
00
830.3
7(0
.000)
⊿L
EV
ER
AG
E132.4
28
47.3
00
2.8
00
0.0
05
485.5
3(0
.000)
126.5
37
46.2
20
2.7
40
0.0
06
426.6
6(0
.000)
⊿V
OL
AT
ILIT
Y19.7
90
10.3
42
1.9
10
0.0
56
165.8
1(0
.000)
22.2
28
11.2
54
1.9
80
0.0
48
171.7
(0.0
00)
⊿R
F-1
06.8
02
140.2
11
-0.7
60
0.4
46
33.3
9(0
.8779)
-47.2
87
117.9
62
-0.4
00
0.6
89
22.7
5(0
.9967)
⊿S
QR
(RF
)(i
i)-3
2.9
19
17.5
64
-1.8
70
0.0
61
19.3
2(
0.9
996)
-46.7
29
16.5
02
-2.8
30
0.0
05
19.9
4(0
.9993)
⊿S
WA
PT
ION
0.0
08
0.0
10
0.7
90
0.4
30
35.5
0(0
.8157)
-0.0
03
0.0
10
-0.3
30
0.7
41
36.9
3(0
.7660)
⊿V
IX-
--
--
0.1
90
0.0
46
4.1
50
0.0
00
194.2
(0.0
00)
log
lik
elih
oo
d
ad
j-R
2
ob
s
0.0
70.2
1
60,5
24
56,5
64
PM
GP
MG
-145,6
27
-132,7
48
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 32 -
Table7 (continued)
Pan
el B
: d
ep
en
den
t v
ari
ab
le=
⊿ln
(CD
SS
)
co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)
lon
g-t
erm
eq
uati
on
LE
VE
RA
GE
6.5
81
0.9
73
6.7
60
0.0
00
7.7
31
0.9
87.9
10
0.0
00
VO
LA
TIL
ITY
2.7
48
0.1
74
16
0.0
00
2.9
48
0.1
716.8
70
0.0
00
RF
-2,8
21
167
-17
0.0
00
-2,6
14
165
-15.8
50
0.0
00
SQ
R(R
F)
(ii)
9.2
69
0.5
37
17
0.0
00
8.6
80.5
316.3
20
0.0
00
TS
-322
15
-22
0.0
00
-297.9
14.3
3-2
0.7
80
0.0
00
TO
PIX
-49
2.5
33
-19
0.0
00
-28.5
2.0
8-1
3.6
60
0.0
00
ln(M
V)
-0.1
60
0.1
51
-1.0
60
0.2
87
-0.0
82
0.1
5-0
.540
0.5
90
SW
AP
TIO
N0.0
52
0.0
03
17
0.0
00
0.0
53
0.0
017.1
40
0.0
00
sh
ort
-term
eq
uati
on
ad
just
men
t p
ara
mer
ter
(ϕi)
-0.0
06
0.0
00
-26.1
50
0.0
00
117.8
7(0
.0000)
-0.0
06
0.0
0-2
3.2
90
0.0
00
110.2
2(0
.0000)
⊿L
EV
ER
AG
E0.6
57
0.1
75
3.7
50
0.0
00
324.5
9(0
.0000)
0.5
94
0.1
54.0
70
0.0
00
284.5
1(0
.0000)
⊿V
OL
AT
ILIT
Y0.2
13
0.0
25
8.5
70
0.0
00
157.6
4(0
.0000)
0.1
78
0.0
27.4
00
0.0
00
129.3
9(0
.0000)
⊿R
F-2
.488
0.4
47
-5.5
70
0.0
00
34.6
2(0
.8436)
-2.9
94
0.4
4-6
.880
0.0
00
34.0
4(0
.8602)
⊿S
QR
(RF
)(i
i)0.1
19
0.0
78
1.5
30
0.1
25
27.1
1(0
.9787)
-0.0
50.0
8-0
.670
0.5
02
27.1
3(0
.9786)
⊿S
WA
PT
ION
0.0
00
0.0
00
4.3
50
0.0
00
20.7
4(0
.9989)
0.0
00
0.0
02.2
30
0.0
26
23.4
4(0
.9953)
⊿V
IX-
--
--
0.0
01
0.0
021.3
90
0.0
00
50.0
4(0
.2461)
log
lik
elih
oo
d
ad
j-R
2
ob
s
-0.0
10.1
3
60,5
24
56,5
64
PM
GP
MG
129,0
10
122,8
39
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 33 -
Table 8: PMG Model (DD, Full Sample)
χ2 is t
he t
est
sta
tisti
cs f
or
the n
ull h
yp
oth
esis
th
at
para
mete
rs a
re h
om
og
en
eo
us, an
d f
igu
res in
pare
nth
eses a
re a
sso
cia
ted
p-v
alu
e. T
he c
oef
an
d s
.e. fo
r S
QR
(RF
) an
d ⊿
SQ
R(R
F)
are
exp
ressed
as d
ivid
ed
by
10000. F
igu
res in
lo
ng
-ru
n e
qu
ilib
riu
m e
qu
ati
on
are
resu
lts a
bo
ut
ho
mo
gen
eo
us p
ara
mete
rs. c
oef
of
oth
er
exp
lan
ato
ry v
ari
ab
les a
re c
ross-s
ecti
on
al
av
era
ge o
f co
eff
icie
nts
fo
r each
fir
ms. s
.e., z
-valu
e, an
d p
-valu
e o
f th
ese v
ari
ab
les a
re t
he s
tan
dard
err
or,
z-v
alu
es a
nd
p-v
alu
es r
ela
tin
g t
o t
he c
on
sta
nt
term
in
th
e O
LS
esti
mati
on
in w
hic
h t
he p
ara
mete
r esti
mate
s o
f each
issu
e a
re t
he d
ep
en
den
t v
ari
ab
le w
hile t
he c
on
sta
nt
term
is e
xpla
nato
ry v
ari
ab
le. ad
j-R
2 is c
ross-s
ecti
on
al av
era
ge. T
he r
esu
lts o
n c
on
sta
nt
term
s a
re o
mit
ted
.
Pan
el A
: d
ep
en
den
t v
ari
ab
le=
⊿C
DS
S
co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)
lon
g-t
erm
eq
uati
on
DD
9.3
74
2.4
17
3.8
80
0.0
00
-0.2
20
0.5
82
-0.3
80
0.7
06
SQ
R(R
F)
(ii)
261
30
8.5
60
0.0
00
63
5.3
69
11.7
50
0.0
00
TS
-105,0
78
10,6
93
-9.8
30
0.0
00
-26,7
49
1,5
36
-17.4
10
0.0
00
TO
PIX
-18,7
02
1,9
51
-9.5
90
0.0
00
-3,5
49
224
-15.8
50
0.0
00
ln(M
V)
-224
29
-7.7
90
0.0
00
-75
7.0
94
-10.5
50
0.0
00
SW
AP
TIO
N22
2.3
86
9.0
80
0.0
00
5.5
01
0.3
66
15.0
20
0.0
00
sh
ort
-term
eq
uati
on
ad
just
men
t p
ara
mer
ter(
ϕi)
-0.0
02
0.0
00
-5.0
20.0
0156.6
1(0
.0000)
-0.0
04
0.0
01
-7.8
00.0
0460.0
2(0
.0000)
⊿D
D-0
.352
0.2
14
-1.6
40.1
020.7
8(0
.9989)
-0.6
50
0.4
45
-1.4
60.1
426.8
8(0
.9804)
⊿S
QR
(RF
)(i
i)-5
9.2
29
22.3
78
-2.6
50.0
122.0
7(0
.9977)
-59.3
86
20.8
79
-2.8
40.0
022.9
4(0
.9963)
⊿S
WA
PT
ION
0.0
26
0.0
27
0.9
70.3
338.8
8(0
.6903)
0.0
15
0.0
18
0.8
40.4
045.0
7(0
.4271)
⊿V
IX-
--
--
0.2
57
0.0
66
3.8
90.0
0217.1
9(0
.0000)
log
lik
elih
oo
d
ad
j-R
2
ob
s
0.0
50.2
0
60,5
24
56,5
64
PM
GP
MG
-146,1
43
-133,3
25
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 34 -
Table 8 (continued)
Pan
el B
: d
ep
en
den
t v
ari
ab
le=
⊿ln
(CD
SS
)
co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)
lon
g-t
erm
eq
uati
on
DD
-0.0
52
0.0
09
-5.7
40.0
0-0
.045
0.0
10
-4.4
30.0
0
SQ
R(R
F)
(ii)
0.7
20
0.0
77
9.3
50.0
00.8
98
0.0
87
10.3
00.0
0
TS
-471
25
-18.5
10.0
0-4
74
27
-17.7
20.0
0
TO
PIX
-79
4.5
47
-17.4
80.0
0-5
43.6
19
-14.9
80.0
0
ln(M
V)
-1.4
87
0.0
96
-15.5
30.0
0-1
.659
0.1
01
-16.4
50.0
0
SW
AP
TIO
N0.0
68
0.0
05
14.3
60.0
00.0
73
0.0
05
14.1
20.0
0
sh
ort
-term
eq
uati
on
ad
just
men
t p
ara
mer
ter(
ϕi)
-0.0
05
0.0
00
-37.8
80.0
0103.0
1(0
.0000)
-0.0
04
0.0
00
-31.3
30.0
080.0
8(0
.0007)
⊿D
D-0
.008
0.0
01
-6.9
60.0
0121.9
9(0
.0000)
-0.0
07
0.0
01
-6.1
00.0
0108.2
9(0
.0000)
⊿S
QR
(RF
)(i
i)-0
.008
0.0
82
-0.1
00.9
228.7
5(0
.9633)
-0.1
62
0.0
83
-1.9
40.0
529.7
2(
0.9
510)
⊿S
WA
PT
ION
0.0
00
0.0
00
5.2
60.0
020.7
1(0
.9989)
0.0
00
0.0
00
2.7
10.0
124.7
6(
0.9
915)
⊿V
IX-
--
--
0.0
01
0.0
00
19.9
40.0
057.3
7(0
.0851)
log
lik
elih
oo
d
ad
j-R
2
ob
s
-0.0
10.1
4
60,5
24
56,5
64
PM
GP
MG
128,5
80
122,4
42
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 35 -
Table 9: PMG Model (Before and After Crisis)
χ2 is t
he t
est
sta
tisti
cs f
or
the n
ull h
yp
oth
esis
th
at
para
mete
rs a
re h
om
og
en
eo
us, an
d f
igu
res in
pare
nth
eses a
re a
sso
cia
ted
p-v
alu
e. T
he c
oef
an
d s
.e. fo
r S
QR
(RF
) an
d ⊿
SQ
R(R
F)
are
exp
ressed
as d
ivid
ed
by
10000. F
igu
res in
lo
ng
-ru
n e
qu
ilib
riu
m e
qu
ati
on
are
resu
lts a
bo
ut
ho
mo
gen
eo
us p
ara
mete
rs. c
oef
of
oth
er
exp
lan
ato
ry v
ari
ab
les a
re c
ross-s
ecti
on
al
av
era
ge o
f co
eff
icie
nts
fo
r each
fir
ms. s
.e., z
-valu
e, an
d p
-valu
e o
f th
ese v
ari
ab
les a
re t
he s
tan
dard
err
or,
z-v
alu
es a
nd
p-v
alu
es r
ela
tin
g t
o t
he c
on
sta
nt
term
in
th
e O
LS
esti
mati
on
in w
hic
h t
he p
ara
mete
r esti
mate
s o
f each
issu
e a
re t
he d
ep
en
den
t v
ari
ab
le w
hile t
he c
on
sta
nt
term
is e
xpla
nato
ry v
ari
ab
le. ad
j-R
2 is c
ross-s
ecti
on
al av
era
ge. T
he r
esu
lts o
n c
on
sta
nt
term
s a
re o
mit
ted
.
Pan
el A
: d
ep
en
den
t v
ari
ab
le=
⊿C
DS
S
co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)
lon
g-t
erm
eq
uati
on
LE
VE
RA
GE
124
28
4.4
50
0.0
00
808
99
8.1
70
0.0
00
VO
LA
TIL
ITY
38
7.1
63
5.3
10
0.0
00
77
13
5.9
00
0.0
00
RF
-4,6
09
2,8
77
-1.6
00
0.1
09
-178,3
31
14,2
28
-12.5
30
0.0
00
SQ
R(R
F)
(ii)
4.4
24
9.3
49
0.4
70
0.6
36
593
46
12.8
80
0.0
00
TS
4,2
38
613
6.9
20
0.0
00
-19,8
25
1,5
44
-12.8
40
0.0
00
TO
PIX
44
46
0.9
50
0.3
41
-1,4
86
128
-11.6
10
0.0
00
ln(M
V)
-6.5
13
3.6
29
-1.7
90
0.0
73
-14
14
-1.0
50
0.2
93
SW
AP
TIO
N-0
.909
0.1
34
-6.8
00
0.0
00
3.0
66
0.2
50
12.2
70
0.0
00
sh
ort
-term
eq
uati
on
ad
just
men
t p
ara
mer
ter(
ϕi)
-0.0
05
0.0
01
-7.9
00
0.0
00
86.9
3(
0.0
001)
-0.0
12
0.0
01
-9.7
00
0.0
00
449.7
0(0
.0000)
⊿L
EV
ER
AG
E35
19
1.8
70
0.0
61
136.6
4(0
.0000)
150
73
2.0
50
0.0
40
211.0
8(0
.0000)
⊿V
OL
AT
ILIT
Y3.3
70
1.3
25
2.5
40
0.0
11
164.6
1(0
.0000)
44
29
1.5
00
0.1
33
161.7
5(0
.0000)
⊿R
F97
90
1.0
80
0.2
81
32.3
3(0
.9035)
-336
242
-1.3
90
0.1
64
24.9
7(
0.9
907)
⊿S
QR
(RF
)(i
i)-8
.202
8.7
97
-0.9
30
0.3
51
26.6
4(0
.9820)
-104
31
-3.3
40
0.0
01
18.9
8(0
.9996)
⊿S
WA
PT
ION
0.0
07
0.0
02
3.2
10
0.0
01
23.3
9(0
.9955)
-0.0
23
0.0
34
-0.6
60
0.5
06
51.0
9(0
.2150)
⊿V
IX0.0
06
0.0
14
0.4
50
0.6
56
31.8
6(
0.9
138)
0.1
82
0.0
49
3.7
30
0.0
00
83.5
5(0
.0003)
log
lik
elih
oo
d
ad
j-R
2
ob
s
Befo
re t
he c
risis
(P
MG
)A
fter
the c
risis
(P
MG
)
-18,5
02
-63,3
41
0.0
70.2
4
33,3
90
23,1
74
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 36 -
Table 9 (continued)
Pan
el B
: d
ep
en
den
t v
ari
ab
le=
⊿ln
(CD
SS
)
co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)co
ef
s.e
.z-
valu
ep
-valu
eχ
2(p
-valu
e)(i
)
lon
g-t
erm
eq
uati
on
LE
VE
RA
GE
2.8
17
2.5
91
1.0
90
0.2
77
2.2
51
0.8
32
2.7
10
0.0
07
VO
LA
TIL
ITY
0.2
82
0.4
85
0.5
80
0.5
61
0.4
39
0.1
40
3.1
30
0.0
02
RF
-353
257
-1.3
70
0.1
70
-2,9
02
161
-18
0.0
00
SQ
R(R
F)
(ii)
0.1
82
0.8
53
0.2
10
0.8
31
9.2
64
0.5
25
18
0.0
00
TS
467
75
6.1
90
0.0
00
-145
16
-8.8
90
0.0
00
TO
PIX
3.3
59
4.1
29
0.8
10
0.4
16
-15
1.2
59
-12
0.0
00
ln(M
V)
-1.0
52
0.3
48
-3.0
30
0.0
02
-0.9
55
0.1
52
-60.0
00
SW
AP
TIO
N-0
.082
0.0
15
-5.3
80
0.0
00
0.0
52
0.0
02
22
0.0
00
sh
ort
-term
eq
uati
on
adju
stm
ent
para
mer
ter(
ϕi)
-0.0
03
0.0
00
-16
0.0
00
22.3
5(0
.9973)
-0.0
16
0.0
01
-23
0.0
00
147.1
7(0
.0000)
⊿L
EV
ER
AG
E0.8
34
0.1
46
5.7
00
0.0
00
184.6
1(0
.0000)
0.2
65
0.1
49
1.7
80
0.0
76
100.8
6(0
.0000)
⊿V
OL
AT
ILIT
Y0.2
47
0.0
32
7.6
00
0.0
00
185.6
0(0
.0000)
0.1
20
0.0
33
3.6
00
0.0
00
95.9
4(0
.0000)
⊿R
F0.8
90
0.4
39
2.0
30
0.0
43
29.2
9(0
.9567)
-8.7
94
0.8
39
-10
0.0
00
46.6
0(0
.3661)
⊿S
QR
(RF
)(i
i)-0
.001
0.0
96
-0.0
10
0.9
94
29.4
9(0
.9542)
-0.4
21
0.1
24
-3.4
00
0.0
01
28.2
0(0
.9692)
⊿S
WA
PT
ION
0.0
00
0.0
00
7.0
00
0.0
00
21.5
5(0
.9982)
0.0
00
0.0
00
-0.6
10
0.5
44
40.5
4(0
.6208)
⊿V
IX0.0
01
0.0
00
6.6
60
0.0
00
34.2
9(0
.8533)
0.0
01
0.0
00
19
0.0
00
25.2
5(0
.9896)
log
lik
elih
oo
d
ad
j-R
2
ob
s33,3
90
23,1
74
83,8
46
45,2
68
0.0
50.1
2
Befo
re t
he c
risis
(P
MG
)A
fter
the c
risis
(P
MG
)
<FSA Institute Discussion Paper Series DP2011-2 (2, 2011)>
- 37 -
Table 10: Principal Component Analysis
Figures are the contributions of principal components.
Before the Crisis After the crisis Before the Crisis After the crisis
1st principa component 38.0% 45.8% 38.5% 38.9%
2nd principal component 6.3% 7.8% 6.2% 7.3%
3rd principal component 5.2% 5.4% 4.8% 6.4%
4th principal component 4.0% 3.9% 4.0% 4.7%
5th principal component 3.6% 3.3% 3.6% 3.6%
Before the Crisis After the crisis Before the Crisis After the crisis
1st principa component 35.5% 58.4% 36.0% 50.2%
2nd principal component 6.9% 5.0% 6.8% 5.5%
3rd principal component 5.1% 4.3% 4.7% 5.0%
4th principal component 4.0% 2.8% 4.0% 3.4%
5th principal component 3.4% 2.4% 3.4% 3.0%
⊿ ln(CDSS)
raw variable residuals from full model
⊿ CDSS
raw variable residuals from full model
Figure 1: Scatter Plot of Eigenvectors of 1st Principal Components
0.00
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
0.20
0.00 0.05 0.10 0.15 0.20
before the crisis
after the crisis
(eigenvector of 1st principal component, using ⊿ln(CDSS)
(eig
en
vecto
r of
1st p
rincip
al c
om
po
nen
t, usin
gre
sidu
al o
f full m
od
el)
Financial Research Center (FSA Institute)
Financial Services Agency
Government of Japan
3-2-1 Kasumigaseki, Chiyoda-ku, Tokyo 100-8967, Japan
TEL:03-3506-6000(ext 3293)
FAX:03-3506-6716
URL:http://www.fsa.go.jp/frtc/english/index.html